Applied AI for Biotech R&D and Biomanufacturing — Upskilling Plan v4

Hours: 8-10 h/week as in v3. If you can do more, the calendar compresses; it never expands in scope.

Most biotech startups do not have an AI problem. They have a data-to-decision problem. Their experimental loop (design → run → measure → analyze → decide → next run) takes weeks because the data lives in spreadsheets, instrument exports, paper batch records and PDFs. The AI that matters to them is not a chatbot. It is the layer that closes that loop: structured capture, models that work on small and expensive data, and a decision every week instead of every quarter. I build that layer, measure how many days it takes out of the loop, and publish the number.

Why this is your niche and not someone else's:

1. It is the same talent you showed at EY. You made delivery visible and measurable when nobody asked. The lab equivalent is "how many days from question to decision, and where do those days go". Nobody in a wet lab has that number.
2. The evidence is in the list. 64 of 86 GridX startups have no AI/IoT base. Their stated gaps cluster into four repeated patterns (section 3), and none of the four needs a new algorithm. They need an engineer who understands the science, the software and the money.
3. The money is real and it is not price-sensitive. Pharma, CDMOs, agbio groups and funded biotech pay for speed (startups: runway) and for risk reduction (pharma: lost batches, audit findings). What your GridX contact told you ("if they are not burning, they are not growing") means: sell weeks saved per experiment cycle, converted into runway, not "cost savings". Startups buy time. Pharma buys risk.
4. Biology sets the pace, not a partner. A fermentation takes 7 days. A field trial takes a season. A stability study takes months. That timescale structurally prevents the "deliver more with less by Friday" culture you are leaving. Scientist-led companies are, on average, the calmest place a senior engineer can work.
5. The AI-native biotechs are your peers, not your clients. Tesabio.ai, Zavia, Nanogrow, CellCo, Dogma, Mavios and the other 16 "AI: yes" companies show what the sophisticated end of the market looks like. You learn from them, cite them, ask them to review or co-author, and sell to the 64 that are ten years behind them in data maturity.

The metric that defines the brand: loop cycle time (days from hypothesis to decision) and data readiness (% of experimental data that is machine-readable within 24 h of being generated). Every instrument you build moves one of the two. Every offer you sell measures one of the two first.

flowchart LR
  T[Theory: bioprocess, DoE, BO, MSPC, hybrid models, doc AI] --> I1
  T --> I2
  T --> I3
  I1[Instrument 1: batchlens\nfermentation batch intelligence] --> E[Evidence: repos, DOIs, 2 preprints, 4 deep dives, 3 videos]
  I2[Instrument 2: trialkit\ntrials and assays to decision] --> E
  I3[Instrument 3: labdocs\non-prem lab document intelligence] --> E
  V[Vara ideas: telemetry schema + calibrated judge] --> I1
  V --> I3
  E --> A1[Audience 1: 64 non-AI GridX startups\nfeedback, data, co-authors, first paid audits]
  E --> A2[Audience 2: established pharma, agbio, CDMOs, funds\nhigh-ticket audits and pilots]
  E --> A3[Audience 3: scientific-data companies\nemployee track, remote, relaxed]
  A1 --> O[Offer: Loop Audit, 2 weeks, fixed price]
  A2 --> O
  O --> P[Pilot 6-8 weeks] --> R[Retainer: fractional AI lead for R&D]
  A3 --> J[Job offers as fallback or by choice]

The chain in words:

• Theory feeds instruments. Every lecture in section 6 exists because a practice in section 7 needs it. Nothing is studied "for later".
• Instruments produce evidence, not products. Each instrument is used on public data first, then on one founder's sample data for free, then inside a paid audit. You never sell the instrument. You sell the number it produces about their loop.
• Evidence opens three audiences at once with the same artifacts. The startup list is your proving ground and your first invoices. Established companies and funds are the high ticket. Scientific-data companies (TetraScience, Uncountable, DataHow, Benchling and the like) are the employee track if consulting is slower than your patience.
• The offer is fixed-price and small on purpose. A 2-week Loop Audit is below most purchase thresholds and can be paid by card or a single invoice. It exists to train the muscle you have never used, and to generate the anonymized case study that sells the next one.
• Funds are the channel. GridX, SF500, CITES and similar funds each sit on 30-100 portfolio companies with the same four gaps. One "portfolio data-readiness package" sold to a fund puts you in front of ten startups without ten sales conversations. This is the "facilitator channel" from your August synthesis, applied to biotech.
• The Vara work is not lost. Its telemetry schema (cost × model × task) and its calibrated LLM-judge method are exactly what makes instruments 1 and 3 credible: measured, not claimed.

3.1 The four gap archetypes in the GridX list

I re-read all 86 descriptions and gaps and grouped them. Counts are companies whose primary gap falls in the archetype (full mapping in Appendix A).

Off-thesis: Cryosmetics (e-commerce recommender). Ignore.

Rules for using this list in public:

• Never name a company's weakness in public. The CSV's "falencias" column is your private hypothesis, not a fact. In posts, talk about archetypes ("a fungal pigment fermentation startup") and name companies only when praising them or after they engaged.
• Tier the 64 non-AI companies by what you need from them, not by what they can pay: (1) five feedback partners who will look at an instrument for 30 minutes, (2) two data partners who will share a sample dataset under NDA for a free run, (3) one co-author or arXiv endorser (most founders are CONICET scientists with publication records).

3.2 Beyond the list: who actually pays high ticket

I have not verified the current status, funding or openings of these companies. Treat this as a research list; each name needs 10 minutes of verification before outreach.

Established Argentine and regional companies (money, regulated processes, slow data):

Funds and programs as channels: GridX (you already have a contact), SF500 (Rosario), CITES (Sunchales), Draper Cygnus, Kamay Ventures (food/agri corporate VC). Offer: "portfolio data-readiness package", 5 startups × 1-week mini audit, one consolidated report for the fund's partners.

Scientific-data and bioprocess-AI companies (inspiration now, employer track later, verify each): TetraScience, Uncountable, Benchling, Scispot, Sapio Sciences, LabKey, Synthace, DataHow (Switzerland, hybrid modeling for bioprocess), Novasign (Austria), Basetwo AI, Aizon (Barcelona, AI for pharma manufacturing, Spanish-speaking), Invert, Quartic.ai, Culture Biosciences, Ginkgo Bioworks, Cradle, LatchBio, Sartorius/Umetrics (SIMCA), Cytiva, Novonesis, DSM-Firmenich, Lonza, Novo Nordisk, NNE. These are the people who will read your preprints. Instrument 1 is, on purpose, a small open version of what DataHow and Aizon sell.

Not on the list on purpose: Big Four and large IT consultancies' life-sciences practices. Same culture you are leaving, different logo.

3.3 Culture filter (use it on every client and employer)

Ask, before signing anything:

1. Who decides priorities: scientists, or a delivery office?
2. What happened the last time a milestone slipped? (Listen for "we re-planned" versus "we worked weekends".)
3. How long is one experimental cycle here? (If the answer is in days, the company is paced by biology. Good.)
4. Is there on-call, and for what?
5. How many people left in the last 12 months, and why?
6. Can I see the last two retrospectives or lab meeting notes?

Two bad answers out of six: walk away. Your friend at Sumble is your calibration point: ask them to answer these six about their own workplace so you know what "good" sounds like.

These prices are hypotheses to test with the first five conversations, not facts.

What the "burn to grow" comment implies for the pitch: never say "save money". Say "one fewer week per cycle × N cycles a year = M months of runway you did not have to raise". For pharma, say "batches not lost, findings not written".

Fixed price, half upfront, paid by invoice or card. No hourly billing in the first year: hourly reproduces EY's logic inside your own business.

Each instrument follows the same seven steps, and the steps are the cure for the "instrument trap":

1. Build on public data.
2. Publish the number it produces (a deep dive, a DOI, a preprint).
3. Run it for free on one founder's sample (30 minutes of their time, NDA if needed).
4. Publish the anonymized before/after.
5. Put it inside a priced Loop Audit.
6. Ask the client for a quote you can publish.
7. Move on. Never add a feature that no client asked for.

5.1 Instrument 1 · batchlens — fermentation batch intelligence (weeks 1-8)

• Gap it closes (archetype A): paper or spreadsheet batch records, no golden batch, deviations detected at harvest, yield optimized by intuition.
• Public data: IndPenSim (industrial-scale penicillin fermentation simulator; the public release has 100 batches with online process variables and Raman spectra, including faulty batches). Verify the download link on industrialpenicillinsimulation.com. Plus your own Monod fed-batch simulator streamed over MQTT so the IoT gap is demonstrable without hardware.
• What it does: ingest batch CSVs or extracted paper records → batch unfolding and alignment → golden-batch envelope → multivariate monitoring (PCA, Hotelling T², SPE) → soft sensor for biomass and product titer from online variables → early deviation alarms → Bayesian optimization suggestion for the next run's feed profile → one HTML report per batch. --fake mode runs everything with no data.
• Stack: Python, pandas, scikit-learn, LightGBM, statsmodels, BoTorch (CPU is fine), Plotly, Mosquitto + Node-RED + InfluxDB + Grafana on a Debian box, optional ESP32 + temperature probe (USD 15) so one real sensor exists in the demo.
• Hardware fit: all models train on CPU or in seconds on the RTX 3050. The 6 GB card matters only in week 5 (vision model for batch-record extraction).
• Headline numbers you publish: hours-to-detect a deviation versus harvest time; titer prediction error at 50% of batch time; simulated titer gain after 12 BO-suggested runs versus a DoE grid.
• Companies to show it to: Michroma, Levya, Fungilife, NanoInGreen, POLYMERA, Hiamet, Food4You. Peers to ask for review: Mavios (they sell the enterprise version of this), Stämm.
• Preprint 1 (week 8): "Early deviation detection and soft sensing on an open industrial fermentation benchmark with consumer hardware: a reproducible baseline". arXiv cs.LG or q-bio.QM (needs an endorser), or bioRxiv (no endorser needed). Zenodo DOI for the code.

5.2 Instrument 2 · trialkit — from trials and assays to a defensible decision (weeks 9-13)

• Gap it closes (archetypes B and C): field trials and plate assays scattered across spreadsheets, subjective scoring, reports built by hand, no confidence intervals, so investors and growers do not trust the efficacy claim.
• Public data: field module: agridat (hundreds of published agricultural trials, CSVs available through the Rdatasets mirror) and Genomes to Fields; plate module: published growth-curve and dose-response datasets, PlantVillage for image scoring; Zindi/CGIAR agro challenges for extra practice.
• What it does: LLM-assisted mapping of any spreadsheet into one trial schema (plot, block, treatment, replicate, date, variable) → design detection (RCBD, split-plot) → mixed model with treatment effects and intervals → plate module: growth curves (logistic, Gompertz), 4PL dose-response, MIC, Michaelis-Menten, Z′-factor QC → image scoring (zero-shot CLIP, then a small fine-tuned classifier) → one report in Spanish or English that a grower, a fund or a regulator can read.
• Stack: pandas, statsmodels (MixedLM, ANOVA), scipy curve fitting, PyTorch small CNN or CLIP on the 3050, Claude or a local Qwen for schema mapping, the Vara harness to measure mapping accuracy and cost.
• Headline numbers: days from last measurement to report (manual versus trialkit); mapping accuracy of spreadsheets to schema; agreement between image scoring and human scoring.
• Companies to show it to: Puna Bio, Nunatak, Bsafe, Semion (field); Resistia, APEXzymes, Antarka, Syocin (plate). Peers: Beeflow, Zavia.
• Output (week 13): Medium deep dive 2, Zenodo DOI, 60-s video, Loop Audit offer launched.

5.3 Instrument 3 · labdocs — on-prem document intelligence with an audit trail (weeks 14-18)

• Gap it closes (archetype D, and every archetype's "records on paper" line): SOPs, protocols, batch records and notebook pages that are not searchable, not traceable and not auditable; agencies (ANMAT, FDA, EMA) demand ALCOA+ traceability the startup does not have.
• Public data: synthetic batch records and SOPs you generate from real templates (never client data), plus public regulatory documents. eLabFTW (open-source electronic lab notebook) running on your Debian box as the substrate, so the demo is "AI inside a real ELN", not a toy.
• What it does: extract structured experiments from PDFs and scanned pages with a local vision model (Qwen2.5-VL 3B or Gemma 3 4B on the 3050) → write to eLabFTW through its API with a hash-chained audit trail → local search and Q&A with citations (embeddings on the Debian box) → deviation report drafting from batchlens alarms plus the relevant SOP, human-in-the-loop → calibrated evaluation: 60 human-labeled cases, judge-human agreement published, local versus cloud accuracy and cost table.
• Stack: Ollama or llama.cpp (Qwen2.5 7B Q4, Qwen2.5-VL 3B, Gemma 3 4B, bge-m3 or nomic embeddings), eLabFTW in Docker, Postgres, the Vara judge, Claude and Gemini only as the cloud comparators.
• Hardware fit: this is where the small boxes become the argument. Publish the table: what runs on a USD 300 mini PC, what needs the 3050, what needs the cloud, with accuracy and cost per 1,000 pages.
• Headline numbers: extraction accuracy per field (local versus cloud); judge-human agreement; minutes from scanned page to auditable record.
• Companies to show it to: Eywa Biotech, AVaTAR MedTech, NanoproX, Synthetic Ocular Restoration, SphereBio, Neocell, plus any pharma QA manager.
• Preprint 2 (week 18): "On-premise document intelligence for GxP laboratories: a calibrated evaluation of small vision-language models on synthetic batch records and SOPs". arXiv cs.CL or cs.DL, or a JOSS submission for the software.

5.4 Phase-2 backlog (do not start before week 20)

• Instrument 4 · clinical report automation (archetype E: Enteria, Microgenesis, PreGen, CASPR, Ubique). Higher ticket, regulated (HIPAA/GDPR/Ley 25.326), needs a partner clinic. Start only if a diagnostic company asks after seeing instrument 3.
• Scripts-to-product engineering for archetype F companies with software gaps (Geoprot, ArgenTag, Monte Caldera, Zavia). Off-thesis but lucrative; take one only as a paid engagement, never as a portfolio project.

All free. Roughly 40 hours over 20 weeks, two 1.5-h blocks per week. Each item names the practice it feeds.

Deferred, and why: Karpathy and CS336 internals (you learn them when an instrument needs them, not before), MIT 6.874 "Deep Learning in the Life Sciences" (excellent, 3 lectures allowed inside the literacy thread), full DoE or biostatistics courses (the NIST handbook covers what you need), protein design (Zavia and Nanogrow's territory; you are not competing there).

Rhythm (same as v3, it fits your week): Tue and Thu theory (1.5 h each), Mon/Wed/Fri code (1.5 h each), Sat code (1.5 h, first to fall in a bad week), Sun publish (1 h). The Thursday block ends with 30 minutes of the science literacy thread or reading one founder's paper.

Legend: G = GitHub, L-Sun = LinkedIn Sunday build log, L-Wed = LinkedIn Wednesday gap note (archetype, never a named weakness), M = Medium, X = arXiv/bioRxiv, Z = Zenodo DOI, V = 60-s video.

7.1 Phase 0 · Positioning (week 0)

7.2 Phase 1 · Instrument 1, batchlens (weeks 1-8)

7.3 Phase 2 · Instrument 2, trialkit, and the first invoice (weeks 9-13)

Gate at week 13: if fewer than 3 calls are booked, week 14 becomes a second selling week and instrument 3 shifts by one week. The calendar bends; the gate does not.

7.4 Phase 3 · Instrument 3, labdocs (weeks 14-18)

7.5 Phase 4 · Consolidate and go outbound (weeks 19-20)

You cannot sell to scientists without reading their language. This thread is small, weekly, and produces one Wednesday gap note per month.

Four notebooks in 20 weeks. They are literacy, not instruments: they never get a release, a DOI or a feature.

9.1 Channels and cadence

Rules:

• The Wednesday note is derived from the week's work, 10 minutes, two paragraphs. If it takes 30, skip it.
• Every post has one number or one admitted failure. Never "the ultimate guide".
• Archetypes, never named weaknesses. Name a company only to praise or after they engaged.
• Zero EY data, names, clients or comparisons. All code on your time and hardware, commits dated.

9.2 Template A · Sunday build log (20 minutes)

Week [N] of 20, building open instruments for wet labs at 1-2 h/day.

This week: [ONE thing built or measured]
The number: [the measurement, with N and an interval when you have one]
Where it fails: [one honest line]
Next: [one thing]

[image: chart, table or terminal capture]
Repo: [link]

9.3 Template E · Wednesday gap note (10 minutes)

A [archetype description: "fungal fermentation startup", "bioinput company running field trials"] typically [the gap, stated as a process fact, not a fault].

The cost is not the spreadsheet. It is [the days lost, the batch lost, the claim an investor does not believe].

An open way to close it: [what the instrument does, one sentence] → [link to the week's commit or chart].

If this is your loop, I would like to hear how you do it today. No pitch.

9.4 Template B · Deep-dive and preprint titles that work

1. "Early deviation detection on 100 open fermentation batches: [X] hours before harvest, [Y]% false alarms"
2. "How early can you predict titer? Soft-sensor error at 25/50/75% of batch time on IndPenSim"
3. "Paper batch records to JSON: a 3B local model versus the cloud, per-field accuracy and cost per 1,000 pages"
4. "From spreadsheet to defensible efficacy claim: [N] published field trials, one schema, one report"
5. "On-premise document intelligence for GxP labs: judge-human agreement [X]% (n=60)"

Rule: the title carries a measured number or an admitted failure.

Preprint skeleton (both preprints): problem in days and money → data and why it is public → method, reproducible in one command → results with intervals → where it fails → limitations of the judge or the metric → what you did not test. Mark every unsupported claim [VERIFY] before submission.

9.5 Template C · 60-second video

0-5 s: the problem, in days or in a lost batch.
5-15 s: one command running. Camera on the terminal.
15-45 s: the chart appearing. No cuts.
45-55 s: where it fails. One sentence.
55-60 s: the link.
No intro, no music, large terminal font.

9.6 Profiles (week 0)

LinkedIn headline: Applied AI for biotech R&D and biomanufacturing | I turn lab and plant data into weekly decisions | Open instruments: batchlens · trialkit · labdocs | 25 yrs leading delivery, ex-Big Four AI transformation lead

About (5 lines): the thesis in one sentence; the three instruments with one number each as they appear; the offer in one line; "everything is open, measured, and reproducible on a USD 300 computer"; link to the repo and the calendar.

GitHub: real photo, one-line bio with the thesis, labloop pinned, profile README with the three numbers.

9.7 Template F · Loop Audit one-pager (week 12)

LOOP AUDIT — 2 weeks, fixed price

What you get
- A map of your experimental loop: every step from hypothesis to decision, with days per step.
- Your two numbers: loop cycle time and data readiness (% machine-readable within 24 h).
- Inventory of every data source (instruments, spreadsheets, paper, ELN) and where the days go.
- Three automation candidates ranked by days saved per cycle, with effort estimates.
- One working prototype on a sample of your data, running on your machine or mine.
- A 45-minute readout with your team.

What I need from you
- 3 hours of interviews (founder, lab lead, whoever keeps the spreadsheet).
- A sample of real data under NDA, or synthetic data if you prefer.

Price: USD [X], half upfront. Case-study price for seed-stage companies in exchange for an anonymized write-up.

9.8 Template D · Targeted message (weeks 8, 13, 19)

Hi [NAME],

I read [their paper / a specific thing about their process]. [One sentence showing you understood it.]

I build open instruments for wet-lab data loops. The last one, [instrument], [one number: "detects fermentation deviations X hours before harvest on 100 open batches" / "turns 10 spreadsheet formats into one defensible efficacy report"]. Code, data and a preprint: [link].

Before this I led AI transformation at a Big Four firm, so I care about the invoice as much as the model.

If [archetype gap] is part of your week, I would like 30 minutes to hear how you do it today. If it is useful, I can run it on a sample of your data, no cost, no pitch.

[NAME]

Budget: APIs USD 150-200 over 20 weeks (extraction comparisons in weeks 5 and 15 are the only expensive runs; everything else is --fake or local). Hardware USD 0-15. Everything else is free.

Weekly board (30 seconds on Sunday):

Gates:

Vanity metrics, banned: stars, likes, followers, hours of video, courses completed.

Honest expectation for week 20: evidence, one or two paid audits, a pipeline of conversations, and a clear view of which of the four archetypes pays fastest. The retainer or the job offer is the month 6-9 target.

• Vara as a flagship. Do not start it. Its harness and judge live on inside batchlens (week 5) and labdocs (week 15).
• Instrument 4 (clinical reporting) before week 20, and any regulated clinical data at all in this phase.
• Protein or molecule design. Zavia, Nanogrow, CellCo own that. You do literacy, not competition.
• Building a LIMS/ELN product. eLabFTW exists. You add AI to it; you do not replace it.
• A fourth instrument, a SaaS, a dashboard product, a marketplace. Module 23 in a lab coat.
• Hourly consulting. Fixed price only in year one.
• Any wet-lab work, any hardware beyond one sensor. You are the data-to-decision layer, not the lab.
• Full courses: MIT 6.874 complete, CS336 complete, biostatistics MOOCs, cloud certifications. The catalog in section 6 is all that fits.
• Naming a company's weakness in public. Ever.
• A second project. El Espejo stays at its current low-effort pace and does not count.

• One block, one thing. Theory or code, never both in the same block. Tue/Thu theory, Mon/Wed/Fri/Sat code, Sun publish.
• Code first, theory when the code asks. Try the practice before the lecture; go to the lecture with a concrete question.
• Tomorrow's TODO is the last line of every block, written in the repo.
• Sunday hour is sacred: build log (20 min), board (5 min), NotebookLM gap detector (20 min, carry-over prompt 5), TODO for the week (15 min).
• Publish even when the week was bad. "This week I did not advance; here is what blocked me" is a valid and well-received post.
• Thursday's last 30 minutes go to the literacy thread or to one founder's paper. Reading their paper is the outreach.
• Boredom is a signal. Three bored blocks on a topic means it is not connected to an instrument. Skip it.
• Hygiene while employed: zero EY data or names, own hardware, dated commits, no comparisons with your employer. Your narrative is what you build, never what your firm does wrong.

Tomorrow's first block (60-90 minutes)

1. Create the labloop GitHub org and the batchlens repo with a 10-line README: what it measures, on which public dataset, when the first number ships (week 4).
2. Pull Qwen2.5 7B Q4 and one vision model with Ollama; confirm they load on the 3050.
3. Change the LinkedIn headline (section 9.6).
4. Post 0, five lines: "20 weeks building open instruments for wet-lab data loops, in public, at 1-2 h/day. First instrument: fermentation batch intelligence on 100 open batches. Week 0."
5. Write tomorrow's TODO: "Monod fed-batch simulator, 80 lines, publishing to MQTT."

In 20 weeks nobody in a lab will ask how many lectures you watched. A founder will ask: how many days does your instrument take out of my loop, how do you know, and where does it fail? This document exists so that you can answer with a link, a number and an invoice.

Archetype counts (primary fit): A · Process & telemetry: 25 · B · Field trials: 8 · C · Lab assays & preclinical: 14 · D · Traceability & QMS: 7 · E · Diagnostics & clinical reporting: 11 · F · AI-native / computational: 20 · Off-thesis: 1. AI/IoT base declared: 22 of 86.

The gap column is the private hypothesis from the CSV, shortened. It is never to be quoted in public.

A · Process & telemetry — instrument: 1 · batchlens — role: Proof partner → paying candidate

B · Field trials — instrument: 2 · trialkit (field) — role: Proof partner → paying candidate

C · Lab assays & preclinical — instrument: 2 · trialkit (plate) — role: Proof partner → paying candidate

D · Traceability & QMS — instrument: 3 · labdocs — role: Paying candidate (regulatory pressure)

E · Diagnostics & clinical reporting — instrument: 4 · phase 2 — role: Phase 2

F · AI-native / computational — instrument: none (peer) — role: Peer: review, cite, co-author, endorse

Off-thesis — instrument: none — role: Ignore