The Dispatch
Thursday, October 22, 2026 · Fremont, CA

Software you can actually trust.

The lede

I build AI systems that measure themselves — observability, evals, and gateways that fail loud instead of lying quiet. Right now I'm building Sentinel, a self-hosted LLM gateway with built-in evals. On the side I'm researching hallucination detection in AI-generated radiology reports.

→ spot the hallucination→ case studies↓ resume.pdf

Section 01 · interactive

Spot the hallucination.

This is a short AI-written paragraph about me. A few sentences are quietly false. Click the ones you think an AI made up — then reveal the answer. The site will score you on precision and recall, which is exactly what my research and evals do, automatically.

AI-generated bio · variant A

Section 02 · live demo

Watch a request go through the gateway.

Pick a question. It travels through a scripted mini-Sentinel — auth, routing, model, eval check — and lands as an answer with latency, token count, cost, and an eval score. No jargon.

Sentinel · gateway trace

01
auth
—
02
route
—
03
model
—
04
eval
—
latency
— ms
tokens
—
cost
—
eval
—

response

Pick a question to start a trace.

Scripted answers. Sentinel is in active development.

Section 04 · work

Four projects, four problems with trust.

observe · evaluate · fail loud

01

SignalOps

● shipped

Log observability and alerting platform — rebuilt ingestion on Redis Streams and load-tested to a million documents.

  • Rebuilt ingestion on Redis Streams with consumer groups, stalled-message recovery, and a dead-letter queue — throughput 2,000 → 9,750 logs/sec (4.8×).
  • Three alert rule types that suppress flapping alerts, backed by a Postgres incident store: 0 false positives across 1,200 evaluation ticks.
  • Load-tested 1M documents: 19,810 docs/sec ingest, sub-15 ms p95 search, cursor paging past OpenSearch's 10,000-hit cap.
  • 92 integration tests against live containers in CI on every push.
PythonFastAPIOpenSearchRedis StreamsPostgreSQLDockerAWS Lambda (SAM)GitHub Actions
→ read the case study→ github
02

Autonomous Resume Reviewer Agent

● shipped

A LangGraph agent that tailors resume bullets to a job description in four steps and explains every edit.

  • Four-step agent: parse → rewrite → critique → explain. Every change ships with a reason the user can inspect.
  • 30-case eval set with 15 injected hallucinations. A revised critique prompt caught all 15 (up from 13), with 2 false alarms.
  • pgvector embedding cache skips repeat API calls across runs.
PythonFastAPILangGraphpgvectorNext.jsTypeScript
→ github
03

Sentinel

● building

Self-hosted LLM gateway with built-in evals. Sits between an app and OpenAI/Anthropic, logs every request, and fails a PR when scores drop.

  • Logs every model request with cost, latency, and prompt/response pairs.
  • Runs eval checks on live answers (groundedness, hallucination, policy) and surfaces regressions.
  • GitHub Actions check that fails a PR when eval scores drop against a baseline.
GoPython eval workersClickHousePostgreSQLRedis
→ read the case study→ github
04

Multi-Client Banking System

● live demo

A 20-thread TCP server with race-condition fixes, 16 REST endpoints, and 27 JUnit tests. Three-person team.

  • 20-thread TCP server; race conditions resolved with synchronized access and atomic updates.
  • 16 REST endpoints with role-based access.
  • 27 JUnit tests covering the concurrency and auth paths.
JavaTCP/IPMultithreadingRESTJUnitDocker
→ github→ live demo

more work

  • AI CRM Insight Dashboard — Internal CRM analytics with LLM-written summaries.
  • Ethical Dilemma Simulator — A classroom tool for teaching AI ethics through branching scenarios.

Section 03 · experience · as a CI run

The résumé, re-run.

The same facts a recruiter needs, rendered like a GitHub Actions run. Click a job to expand its logs. Hit re-run to replay from the top.

workflow: career.yml · run #3
President
2024 — present
HackHayward
  • Raised $20,000 from 12 sponsors.
  • Led a 15-person team running a 200-person hackathon with 11 judges and 2 keynote speakers.
  • Helped organize a Google ADK technical workshop and an AI venture-capital fireside chat.
  • Contributed to the development of the HackHayward website.
VP, Technical Events
Aug 2024 — present
Google Developer Groups — CSU East Bay
  • Ran 12 technical events across the chapter year.
  • Co-led a workshop at the GMIS × CAHSI summit in Albuquerque teaching ~225 students to build an AI code review assistant with Google ADK and Gemini.
Learning Assistant
Jan 2023 — Apr 2024
CSUEB STEM Lab
  • Tutored 20–50 students per term across four semesters.
Section 05 · people
Building people, not just code.
I tutored four semesters, co-led a workshop that reached ~225 students in Albuquerque, ran a 200-person hackathon with 11 judges, and raised $20K from 12 sponsors for it. Teaching is how I learn.
0
hackers at HackHayward
$0
sponsor dollars raised
~0
students taught at GMIS × CAHSI
0
GDG events hosted
0
team led
0
semesters tutoring
HackHayward (nonprofit)
President · 2024 — present. I lead a 15-person team running a 200-person hackathon. In 2025 I raised $20,000 from 12 sponsors, helped organize a Google ADK workshop and an AI venture-capital fireside chat, and contributed to the HackHayward website.
Google Developer Groups · CSU East Bay
VP, Technical Events · Aug 2024 — present. Ran 12 events this year and co-led a workshop at the GMIS × CAHSI summit in Albuquerque teaching ~225 students to build an AI code-review assistant with Google ADK and Gemini.
Section 06 · research
Teaching a model to write Findings.
research in progress · preliminary results · advisor Dr. Lynne Grewe
Generating radiology Findings from Impressions on MIMIC-CXR
Fine-tuning a model to generate the detailed radiology Findings for a chest X-ray given the radiologist's original Impression. The goal is a system that can draft clinically faithful Findings — the harder-to-write half of a report — from the shorter conclusion a radiologist already produced.
What's done
● completed
  • Processed 127,650 eligible report pairs from MIMIC-CXR.
  • Fine-tuned Gemini 3.5 Flash on the Impression → Findings task.
  • Evaluated on a 19-report test set vs. the un-tuned baseline.
BLEU
+134%
ROUGE-L
+40%
METEOR
+46%
vs. the un-tuned baseline, on a 19-report test set. Preliminary.
What's planned
▸ planned
  • Give the model the original Impression and its chest X-ray image to generate Findings — dual-modality.
  • Compare dual-modality against the current text-only system to see whether the image improves clinical accuracy and reduces hallucinations.
  • Expand evaluation beyond the current 19-report test set.
synthetic walk-through · not a real patient
① radiologist impression (input): "Atelectasis in the right lower lobe. No pneumothorax."
② model-generated findings: "Low lung volumes. Patchy opacity in the right lower lobe, most consistent with atelectasis. No evidence of pneumothorax or pleural effusion."
→ The current text-only model drafts Findings from the Impression alone. The planned dual-modality version will also receive the chest X-ray and can disagree with the Impression when the image doesn't support it.
No real patient data appears on this site. Any example below is clearly synthetic. Metrics above are preliminary and from a small (19-report) test set.
Section 07 · optional
Prefer a shell?
Type help to list commands. Try whoami, ls projects, cat signalops.md, or sudo hire.
nidhi@dispatch:~$ type `help` to list commands
nidhi@dispatch:~$