🎓 Claremont McKenna College '29 · Computer Science at Harvey Mudd College 🛠️ Backend & infra, retrieval systems, real-time pipelines, sandboxed execution 🔎 Building AI candidate sourcing for an IT staffing firm

📫 rushiljaiswal1@gmail.com  ·  LinkedIn

vaudit: does an RL grader accept work that is correct but different? 🎯 Five CHIP-8 ROMs, each isolating one documented platform quirk, each with its expected frames and falsifiers registered before the bytes existed. Three of five published grading strategies rank a blank screen above a correct implementation, because two real conventions disagree about one instruction and the glyph lands four pixels apart. Both are right. 🧪 The follow-up study falsified four of its own predictions, including my own algebra, and bounded the failure in closed form at φ(1−d) > ½. At realistic GBA density with one displaced sprite the effect disappears, and the write-up says so in those words because I had pre-committed to saying it. 🐛 Seven third-party interpreters as the population, adjudicated by the test ROM's own pass/fail marks rather than by agreement with mine. Six bug reports filed upstream and a pull request that takes one interpreter from 8 failures to 0. Twenty occasions where this project's own measurements were about the wrong thing are logged with traced causes.

Telos: AI candidate sourcing over a 31k-resume corpus 🔎 Hybrid retrieval across four lanes, vector, BM25, keyword, structured filter, feeding a reranker that blends cosine similarity with placement history, capped so candidates with no track record still compete on resume quality alone. 🧪 The result I'm proudest of is a negative one. Six signals all put the recruiter's actual picks past rank 250 of 6,105, so nothing we could extract from a resume predicted recruiter choice. The honest reframing became "2,057 qualified, 70% cold and 30% with prior contact, sort by that" rather than a top-10 we couldn't justify. 🐛 Found a us_only=True default nobody had chosen that was silently discarding 36% of all recruiter submissions, and it did its worst damage on exactly the technical requisitions where retrieval already looked weakest. 📊 Ranker changes score offline against frozen pools and a human-labeled set (BAD@k / recall@k, holdout split), so a change is evaluated without re-calling the LLM.

pipecat-firewall: a drop-in security firewall for Pipecat voice agents 🛡️ Reads each caller turn and blocks, flags, or redacts prompt injection, PII extraction, and policy-override pressure before the LLM runs, a blocked turn never reaches the model, so there's no prompt left to argue with and you don't pay for the call. ⚡ Detectors are deliberately small and deterministic: no API key, no extra model call, microseconds on clean turns. Refusals are canned replies that bypass the LLM entirely. 📦 Pinned <1.0 against the pipecat-ai frame API after testing across 0.0.98–0.0.108, 1.x restructures frames, so the cap is deliberate rather than neglect.

Goodhart: an autoresearch loop that hardens RL verifiers against reward hacking 🎯 A breach is a solution that passes the grader and fails a held-out oracle. A red swarm discovers them, a green team patches the grader, and a regression gate accepts a patch only if the breach now fails and the gold solution still passes, so over-tightening surfaces as friendly fire instead of a win. 🔒 The seed exploit list is empty and an anti-theater check enforces it, so every breach was genuinely discovered. Grader and oracle never share a test case, and no LLM sits in the verdict path, every verdict is a sandboxed subprocess parsed from JUnit XML, never exit codes. 🔧 I built the serving layer: the event bus, the websocket server, and the 2D/3D dashboard that animates off a 7-event stream, one seam that both the live engine and the recorded replay drive, so the demo can't drift from the real run. The golden run is a genuine recording, not a script.

Strata: real-time AI guidance for plumbing contractors in the field 📴 Built for a contractor in a crawl space with no signal: full diagnostic guidance from a local SQLite FTS5 cache offline, with feedback and job summaries queued in an idempotent outbox that reconciles on reconnect. 📷 Camera-first diagnosis through GPT-4o Vision walking a hypothesis-elimination flow, plus hold-to-speak voice and a "Hey Strata" wake word for hands-free use. 🧤 The constraints drove the interface: 56px touch targets for gloved hands, 7:1 contrast for direct sunlight, one clear action per screen.

Vibecheck: plain-English security review for AI-assisted founders, with a one-click fix PR 🔍 Runs Semgrep, Gitleaks, and OSV over a repo, then translates the raw findings into what's wrong, why it matters in real terms, and a copy-paste fix, for founders who shipped something they can't audit themselves. 🔒 The scan engine never executes the target repo: shallow read-only clone, size and file caps, scanners as bounded subprocesses, clone deleted after. 🚀 GitHub OAuth unlocks private repos, Postgres row-level security isolates every scan, and Claude rewrites the files to open the fix PR.

Yaad: a grounded memory engine for dementia care 🧠 An assistant that invents a detail about someone's family is dangerous, so this one refuses when it can't cite a row, and "did I take my pills today?" routes to today's med_logs rather than being answered by nearest-neighbor text. 🎙️ I built the voice agent end to end: LiveKit + Pipecat pipeline, VAD tuning and barge-in, Groq Whisper STT, MiniMax TTS, and a dlib face-match server for the greeting path. 🌏 The Hindi path failed silently until I translated the query to English before the memory lookup, the memories are stored in English, so retrieving in Hindi matched nothing.

_{Also: Sentry, a voice agent that patches its own guardrails mid-call: leak a card number to it and it clusters the failure, has Nemotron write new rules, hot-swaps them from a mutable ActiveConfig, and refuses the same attack seconds later · hindsight-extension, grades your past predictions against what actually happened, then corrects future advice for your blind spots · kalshi-iv-research, Kalshi unemployment markets Granger-cause VIX at a 2-day lead (p=0.024, 222 days), with CPI markets kept as a null control · tab-memory-agent, persistent intent-level memory for your browser · intern-tracker, parses application emails and scores company fit against your real background · signal, AI-curated daily newsletter on a cron}