Day 91 — Where next¶
You finished the ladder. You have a portfolio that spans building AI systems, adapting models, and shipping and operating products on top of them — the horizontal bar of a "T". Day 91 is choosing the vertical to descend into.
There's no single right answer, and there's one honest heuristic for finding yours:
Which week did you not want to end? The ship you kept tinkering with after Sunday, the failure that nagged at you, the topic you opened extra tabs about — that's your tell. Follow it.
Each vertical below is a deferral — something the 90-day program deliberately left out (see the deferred scope and TOOLKIT.md's "not in the core path") so it could keep you shipping. "Next" is just picking one up.
1. Take the big swing pro¶
If your Weeks 9–12 product has real users, the highest-ROI move is often not more learning — it's momentum. Turn it into a side project with revenue, an open-source project with a community, or a startup wedge. You already have the eval suite, the traces, and a launch write-up; keep shipping on a lighter cadence and let it compound.
First moves: talk to ten users; find the one workflow they'd pay for or miss; cut everything else. If it's OSS, write the CONTRIBUTING guide and triage your first outside issue. Keep the eval gauntlet running against each release.
2. Descend into the model layer → Model / Research Engineer¶
What the program skipped: training from scratch, RL (RLHF / DPO at depth / GRPO), reproducing papers, and the math the moment it becomes a real bottleneck.
First moves: work Karpathy's Zero-to-Hero in full and build GPT from scratch; then reproduce one paper end-to-end and publish the repro with an honest "what matched, what didn't." That single artifact is the credential for research-adjacent work.
You're suited to this if the fine-tune week (05) or the internals of the transformer lit you up.
3. Descend into systems → Platform / LLMOps / Infra Engineer¶
What the program skipped: serving internals (vLLM internals, quantization kernels, a real taste of Triton/CUDA), distributed training (FSDP, multi-GPU), and MLOps at platform scale (Kubernetes, Ray, model registries, autoscaling inference).
First moves: serve an open model behind a load-tested API to a real latency/throughput SLO; profile and cut the p99; then take one model from notebook to a monitored, versioned, autoscaled deployment. This is the natural next given the networks/OS/cloud focus — it turns "can ship an LLM app" into "can run inference at scale, on budget."
You're suited to this if the serving/cost/latency numbers — Weeks 6–8 — were the fun part.
4. Specialize in AI security → the sharpest edge for this track¶
Every ship carried a "Secure it" checkpoint; the deep version is a whole discipline — LLM red-teaming, adversarial ML, prompt-injection defense as a system, guardrail engineering, and the OWASP LLM Top 10 as a specialty (garak, PyRIT).
Given a security/infra background, this is the most differentiated vertical on the board: very few people can both build AI systems and break them, and that intersection is exactly where the Plaintext securing-AI / attacking-AI tracks live. Descending here closes the loop between the two curricula.
First moves: run garak against every model you shipped and write up the findings; build a prompt-injection test suite that's part of your eval gauntlet; then red-team someone else's public agent (with permission) and publish the responsible-disclosure write-up.
You're suited to this if the injection and lethal-trifecta threads (Weeks 3–4) were the weeks you dug deepest.
5. Data-centric AI → quieter, consistently in demand¶
What the program skipped: dataset engineering at scale, synthetic-data pipelines, curation and labeling as a craft, and evals-as-a-product.
First moves: build and publish a curated dataset on the Hub with a real datasheet; then a synthetic-data pipeline with validation, and measure whether it moves a downstream eval. Less flashy than the others, reliably employable.
You're suited to this if the dataset-is-the-program lesson (Week 5) or the eval gauntlet (Week 8) is where you found the leverage.
Running underneath all five: convert the portfolio to outcomes¶
This is where "your repos are the credential" cashes in. Don't treat it as separate from the technical path — do it alongside whichever vertical you pick:
- The interview loop is already built. Your ships + the eval gauntlet + the write-ups are the portfolio a strong AI-engineering interview probes. Point people at them.
- OSS → maintainership. The contribution from the big-swing visibility loop can grow into a maintainer role on a tool you actually use — rare, credible signal.
- The build log → an audience. Keep publishing. The dated, measured, honest write-ups are exactly the content the field is short on.
- Keep it fresh. Run the frontier-watch, re-run the gauntlet against each new model, ship something small every few weeks. A cold portfolio reads as a finished hobby; a warm one reads as a practitioner.
The one recommendation, if you want one¶
For someone who came in through the networks / OS / automation / security / cloud door: #4, AI security. It's the one vertical where your existing background is a multiplier instead of a restart, it's genuinely underserved, and it turns two curricula into one story — you can build the system and prove where it breaks.
But trust the heuristic over the recommendation. Reread your Week-1 write-up, notice which week you'd happily do again, and go down that hole.