Bengaluru, India · Backend · Systems · AI Tooling

Systems that don’t drop a line. Tools that tell you why.

I like software that can explain itself. Most systems fail quietly, so I build the parts that notice: pipelines that keep every record, tools that say which stage broke and show their working. Final-year engineering student, IEEE published, happiest a few layers below the UI.

01

Selected work

Python · asyncio · SQLite/FTS5 · TCP · DockerCurrent build

Problem

Grepping a fast-moving log file tells you almost nothing. LogScope follows files live through rotation, groups similar lines into ranked patterns, flags error spikes with statistics it can show you, and searches everything. All of it from the terminal, all on one asyncio event loop.

Design decisions

  • Pattern grouping runs on a hand-built Drain implementation. One bug printing 10,000 near-identical lines becomes one row with a count.
  • Spike detection uses a rolling z-score, and the tool renders the statistic as a small graph. An alert should be an explanation, not a mystery.
  • Every source feeds one bounded queue, drained through parse, index, cluster, anomaly, and sink stages. Bursty sources slow themselves down instead of eating memory.
  • Multi-machine collection runs over a custom length-prefixed TCP protocol with at-least-once delivery. Per-line IDs make ingestion idempotent; agents buffer through disconnects and replay with backoff. Restart the server: no line dropped, none duplicated.
  • An optional AI summary explains the root cause of a problem cluster. It is cached, time-limited, and the tool works fine without it.
  • LogScope watches itself the way it watches your app: events per second, queue depth, lag, query speed.

Outcome

Built to learn how real log tools work on the inside, not to replace Datadog or Loki. Tested like it matters anyway: 100+ tests under GitHub Actions CI, including crash-recovery paths.

github.com/chinmai-sd-123/LogScope ↗

Python SDK · FastAPI · LangGraph · Redis · Next.jsLive, SDK on PyPI

Problem

“The model hallucinated” is not a diagnosis. TraceroAI traces every RAG answer through retrieval, context, and generation, then tells you which stage actually failed.

Design decisions

  • Every trace collapses into one of six named diagnoses, not a vague quality score.
  • Evaluation runs in two tiers. Embedding cosine-similarity checks run on every trace; an optional LLM-as-judge checks claim-level groundedness only where it earns its cost.
  • A LangGraph recovery agent retries the diagnosed stage, not the whole pipeline, and escalates to human review after a bounded number of attempts.
  • An experiment harness A/B-tests retrieval configurations against the same traces.
  • Ingest is multi-tenant behind project API keys. Reads stay open, so you can explore the live data without one.

The six diagnoses

  • healthy_answer
  • correct_refusal
  • retrieval_miss
  • unsupported_claim
  • wrong_answer
  • needs_review

Outcome

Live at traceroai.tech with interactive docs, published to PyPI as traceroai, and written up on DEV.to.

traceroai.tech ↗traceroai.tech/docs ↗pypi.org/project/traceroai ↗github.com/chinmai-sd-123/TraceroAI ↗write-up on dev.to ↗

FastAPI · PostgreSQL · Redis · React + Vite · PrometheusLive

Problem

Resumes claim; repositories prove. SignalStack turns a job description into measurable outcomes, analyzes a candidate’s actual repository code, and produces reports where every score links to a real file, commit, or snippet.

Design decisions

  • Scoring runs in two stages: cheap deterministic screening for everyone, expensive LLM analysis only for candidates who clear it. Roughly 5× fewer LLM requests and 60% lower token cost.
  • Evaluation is fault-tolerant by design. A failed LLM call never overwrites a valid report with a zero; failures are quarantined and retried.
  • The Redis-backed queue survives restarts. Alembic migrations, RBAC, audit logs, and Prometheus-style metrics treat it as a production system, not a demo.

Outcome

Live on Vercel and Render, backed by PostgreSQL on Neon and Redis Cloud, validated across 26 backend test files.

signal-stack-two.vercel.app ↗github.com/chinmai-sd-123/SIgnal_Stack ↗

FastAPI · XGBoost · InfluxDB · Node · ReactPublished, IEEE ICTBIG 2025

Problem

Crew-health telemetry cannot wait for a batch job. Vitals stream in every 10 seconds; the platform classifies each crew member as stable, caution, or critical in real time, against a sub-50 ms inference target.

Design decisions

  • Inference reads a 144-step window (24 minutes of vitals) under a strict feature-order contract, so training and serving can never disagree about the input.
  • XGBoost trained on 49K+ records with 68 engineered features, plus 49K+ TimeGAN-augmented sequences to cover the failure modes real data rarely shows.
  • IsolationForest handles anomalies and Prophet handles forecasting, layered under the classifier rather than competing with it.
  • It is a whole platform, not a notebook: vitals land in InfluxDB, a FastAPI ML service sits behind a Node API gateway, React dashboards render live risk, MongoDB stores users and alerts.

Outcome

93.1% accuracy and 25% better alert reliability than baseline. Published as author at IEEE ICTBIG 2025.

astronaut-space-health.vercel.app ↗github.com/chinmai-sd-123/astronaut_space_health ↗

02

What I can do

  • Concurrency that survives bursts

    LogScope runs on a single asyncio event loop where every source feeds one bounded queue, and multi-machine collection replays over custom TCP with at-least-once, idempotent delivery.

    Python · asyncio · bounded queues · TCP · backpressureLogScope ↑

  • Storage picked to fit the query

    FTS5 full-text search behind a query grammar that compiles to SQL or a live predicate; Postgres with Alembic migrations, Redis for queues and cache, InfluxDB for 10-second telemetry.

    SQLite/FTS5 · PostgreSQL · Redis · InfluxDB · MongoDBLogScope ↑SignalStack ↑Astronaut ↑

  • Making LLM output accountable

    Two-tier evaluation with embedding checks and an LLM-as-judge, a six-diagnosis taxonomy, token and cost accounting, and pipelines where a failed call is quarantined instead of zeroing a valid report.

    LLM-as-judge · embeddings · structured outputs · cost accountingTraceroAI ↑SignalStack ↑

  • Agents and SDKs people can install

    A tracing SDK published to PyPI, an experiment harness for A/B tests, and a LangGraph recovery agent that retries the diagnosed stage with bounded attempts and human escalation.

    Python SDK · PyPI · LangGraph · A/B harnessTraceroAI ↑

  • Real-time ML under a latency budget

    XGBoost on 49K+ records with 68 engineered features and TimeGAN augmentation, reading 144-step windows under a strict feature-order contract against a sub-50 ms target.

    XGBoost · TimeGAN · IsolationForest · ProphetAstronaut ↑

  • Operating it in production

    Prometheus-style metrics, RBAC and audit logs, queues that survive restarts, CI on every push; LogScope even reports its own ingest rate, queue depth, lag, and latency percentiles.

    Prometheus · Docker · GitHub Actions · RBAC · audit logsSignalStack ↑LogScope ↑

  • And the plain list

    Python, C++, JavaScript, TypeScript, SQL; React and Next.js when the system needs a face.

03

Recognition

  • Publication“Design and Development of an AI-Enabled Space Crew Health Monitoring System”IEEE ICTBIG 2025, author
  • CertificationMachine Learning SpecializationAndrew Ng · DeepLearning.AI & Stanford
  • EducationB.E. Information Science, DSCE BengaluruCGPA 8.82 / 10 · 2023–2027
04

Ask the portfolio

A small assistant grounded in this portfolio. Ask it about any build: it streams from /api/assistant, and if the model is unreachable a local index answers instead.

No questions yet. Try one of these:
05

Contact

Hiring for backend, systems, or AI tooling? Write to me. I read everything.

chinmaisdinesh@gmail.com

github.com/chinmai-sd-123 ↗linkedin.com/in/chinmai-sd ↗