$ whoami --deep

What Pulls Me

Not a résumé — that's the uptime --career line. This is the map of what I'm drawn to.

ishan · $ whoami --deep

ishan@prod:~$ cat ~/curiosity

# retrieval
Tree-structured & agentic RAG (PageIndex) — moving past vector-only. Reads more like how you'd actually search a codebase or a textbook. In practice: FTS5 and local embeddings fused by Reciprocal Rank Fusion, because lexical and semantic recall fail on different queries.
↳ rasputin_loop ↳ pdfsage
# agent-tooling
Running agentic coding tools (Conductor, Superset) against real PRs to see which loop converges — Cursor stays the daily driver. The 'agent reads a repo and writes its scaffolding' idea started early.
↳ repomaster
# the-model-layer
Karpathy's nanochat lineage — single-GPU, a reference for what 'tiny but real' looks like down at the model layer.
# agent-boxes
Hermes & OpenClaw on a Hostinger VPS, kept alive to watch where production-grade loops degrade: tool-loop convergence, memory bleed, behaviour under sustained load.

ishan@prod:~$ ls -t ~/obsessions

# agent-memory
Procedural memory (AST-indexed) vs episodic (vector-backed) — the two halves coding agents need. Retrieval precision beats recall when a human gates the library; and a memory without provenance is a rumour, so every episode gets stamped with who, where and when.
↳ skillforge ↳ amadeus ↳ rasputin_loop
# harness-engineering
Wiring evals & guardrails into the LLM workflow itself, not bolted on after. Open question: pre-tool, post-tool, or a separate critic agent? The part I keep landing on: the boundary that stops a plausible skill from activating itself.
↳ amadeus
# multi-agent-orchestration
Agent boundaries that follow output artifacts, not 'roles' — most of the work is contracts between agents and retry semantics.
↳ autodocxpdf ↳ market-research
# production-under-limits
The agentic frontier has to sit on real free-tier limits — I'm drawn to designs where the limit becomes an architecture input, not a wall.
↳ autodocxpdf

ishan@prod:~$ git log --oneline ~/convictions

8007e90 rate-limits-are-architecture
Rate limits are an architecture input, not an ops nuisance — the cheapest request is the one you never make. Reduce calls; don't just pace them.
↳ autodocxpdf
9db45f8 concurrency-is-layered
Three overlapping concurrency controls aren't paranoia — each fails differently (lock TTL expiry vs txn serialization vs lost update). Audit trails are cheap early, impossible to retrofit honestly.
↳ ticketflow ↳ irctc
1320d5d agents-follow-artifacts
Agent boundaries should follow output artifacts (report sections), not 'roles'. BYOK shapes architecture — no server-side key custody simplifies everything.
↳ market-research
782e90f constraint-is-style
No build step forces the discipline frameworks let you skip — separation of content from layout, by hand.
↳ ai-bubble

$ cat ~/certifications

Certifications