A running, dated log of what I'm reading, building, breaking and fixing in AI — not a blog, a lab notebook. AI news, learning notes, experiments, paper notes, project updates, and mistakes & fixes, all in one place.
Total Entries
Logged This Week
Current Streak
Categories Tracked
Consistency across the research log
14 entries found
Swapped pure vector search for a hybrid retriever to see if lexical matching recovers exact-term queries that embeddings miss.
Hypothesis — Hybrid retrieval improves recall on short, keyword-heavy queries without hurting semantic recall.
Outcome — Recall@5 on keyword queries up ~18%. Latency +40ms per query — acceptable for the notes app.
Notes generator was losing generated drafts whenever a widget triggered a rerun mid-generation.
✗ Mistake — Draft text was stored in a local variable instead of st.session_state, so any rerun wiped it.
✓ Fix — Moved generation output into st.session_state and gated the generation button with a guard flag.
Revisited the original RAG paper to sanity-check how the retriever and generator are trained jointly vs. the retrieve-then-read setups I usually build.
Skimmed the release notes for the latest Sonnet model — logging it here since a chunk of the portfolio stack (this Research page included) leans on it.
Shipped separate dashboard views per role instead of one dashboard with conditional rendering — cut load time and simplified permission checks.
BackgroundTasks is fine for fire-and-forget emails, but anything that needs retries, visibility or backpressure wants a real queue (RQ/Celery) — wrote this down before I reach for BackgroundTasks out of habit again.
Tried caching the static system + few-shot portion of the prompt to see how much of the round-trip is prompt processing vs. generation.
Hypothesis — Caching the static prefix should meaningfully cut time-to-first-token on repeated calls.
Outcome — ~35% faster time-to-first-token on cache hits. Worth wiring into the notes generator next.
Load-tested the TMS API and started seeing timeouts once concurrent requests crossed ~40.
✗ Mistake — Each request was opening a fresh session instead of reusing the pooled engine's sessionmaker.
✓ Fix — Centralized session creation through a single sessionmaker and bumped pool_size to match expected concurrency.
Re-read this to double check whether few-shot CoT examples are still worth the token overhead for the resume analyzer's scoring logic.
Logging the timeline since it's a good reminder to always check model availability before shipping a hard dependency on a single model tier.
No more tailwind.config.js — theme tokens now live in @theme blocks inside CSS. Updated my mental model before touching this portfolio's globals.css.
Replaced the old 'coming soon' page with an actual research dashboard — search, filters, timeline view and a learning streak tracker.
Tested 256 / 512 / 1024 token chunks with 10% overlap to see the effect on downstream answer quality.
Hypothesis — 512-token chunks balance context sufficiency against retrieval precision better than 256 or 1024.
Outcome — 512 tokens won on answer relevance; 1024 tokens hurt precision more than expected.
Users bounced between /sign-in and the dashboard in an infinite redirect after a fresh deploy.
✗ Mistake — Production env vars for the Clerk redirect URLs still pointed at localhost.
✓ Fix — Synced NEXT_PUBLIC_CLERK_* redirect URLs with the deployed domain and added a env-var checklist to the deploy script.
