Software Development Engineer, R&D

Ninestars Information Technologies · Bangalore · Jun 2025 – Present

  • Python
  • Django
  • Django Channels
  • Celery
  • RabbitMQ
  • PostgreSQL
  • TimescaleDB
  • Redis
  • Qdrant
  • MinIO
  • AWS
  • Azure
  • Google Cloud
  • GraphQL
  • WebSockets
  • RAG
  • Docker
  • Jenkins

AOTM Vault1

  • Architected and developed AOTM Vault, a multi-tenant secrets management platform supporting 50+ cryptographic algorithms (including post-quantum), multi-cloud storage abstraction (AWS, Azure, Google Cloud), automated Celery-based key rotation policies, real-time WebSocket file streaming, and immutable audit logging across vaults, resources, subscriptions, and billing, reducing enterprise key rotation overhead by 80%.

AOTM FDI8

  • Architected the backend of AOTM FDI, a multi-tenant credential-fraud detection platform for a global credential-evaluation organisation (11 Django apps, 119 REST routes, 43-table PostgreSQL schema): a 5-stage OCR-to-clustering analysis pipeline over RabbitMQ quorum queues, served by five queue-specialised Celery pools behind a uvicorn ASGI app.
  • Made the pipeline safe under at-least-once delivery: a versioned request/reply envelope, per-template fan-out joined by a row-locked fan-in barrier that releases exactly once, idempotent replace-scoped writes, and presigned per-object MinIO URLs so workers never hold storage credentials.
  • Built the rule engine: a depth-bounded boolean DSL (20 typed fields × 14 operators) validated at write time, three-valued evaluation that marks unverifiable inputs unevaluable instead of passing them, ~7 queries per docket regardless of rule count, and Gemini-rephrased reasons verified mechanically before display.
  • Designed the verdict model: weighted rule failures and visual-anomaly severity scored 0–100, blended and banded on tenant-configurable thresholds, and re-derived transactionally as late evidence lands, so verdicts no longer depend on worker timing; separated lifecycle (docket_status) from judgement (final_verdict) via backfill migrations.
  • Modelled versioned master-template generations per institution, credential and award year, matched on perceptual hashes (pHash/dHash) and embeddings of seals, logos and signatures; ranked candidate templates by per-document comparison verdicts and excluded byte-identical copies by SHA-256.
  • Gated evaluation on a completeness check (required documents, fields and a matching master), and built bulk reprocessing as beat-fed, cancellable jobs with dry runs and row locks, where a restart from any stage first proves every earlier stage finished on each document's last run.
  • Designed list APIs on the HTTP QUERY method with multi-value filters and computed-score ranges and sorts in bulk queries, plus WebSocket progress streams, optimistic-locked collaborative editing over Django Channels, and PDF/Excel report generation.
  • Enforced fail-closed multi-tenancy across 33 tenant-scoped tables with ContextVar-scoped managers that raise rather than leak cross-tenant rows, plus RBAC rank guards, rotating JWT sessions and chunked AES-256-GCM export encryption; shipped 170+ versioned builds in seven weeks, backed by 70 scripted test suites.

Search, Ledger and Platform3

  • Built an enterprise RAG system with hybrid vector search (dense+sparse+RRF) over Qdrant, integrated with self-hosted LLMs and Jina Embeddings, enabling sub-800ms semantic search across 100K+ files; and designed a private blockchain ledger with Proof-of-History consensus, immutable transaction chains, Merkle-root block validation, and on-chain document signing/verification for tamper-proof audit trails.
  • Engineered a distributed database migration engine with a long-running daemon process architecture (Unix socket IPC, background thread pools) for agent-side parallel batch extraction and direct-to-S3 upload of 60M+ row databases, with live progress reporting, dynamic thread resizing via backend config sync, and graceful stop, mirroring server-side Celery task behavior entirely on the client side (on-prem server, machines, VPS).
  • Built a multi-tenant rate-limiting + usage-based billing layer with atomic Redis Lua admission (token bucket + per-minute/hour/day fixed-window counters), hierarchical user/project limits, two-phase credit reservation backed by Postgres-mirrored reserved columns + an orphan-reservation sweeper, Celery signal-driven cache invalidation, anniversary-based monthly refills, and partial-processing fallback that auto-trims document batches to the user's affordable cap; sub-5ms hot path with zero Postgres queries on cache hit.

Sensor ML and Telemetry2

  • Fanned each sensor uplink out to an external ML worker pool across 10 RabbitMQ quorum queues (forecast, anomaly, threshold, LLM insight, photogrammetry, CFD), chained on a transactional reply-side fan-in barrier (SELECT FOR UPDATE + on-commit publishing + UUIDv5-derived keys for redelivery idempotency), and integrated CFD and photogrammetry as asynchronous job types with design snapshots frozen under row-level locks, worker-side S3 uploads registered by deterministic file IDs, floor-geometry-to-row placement sync enabling 3D proximity ranking of sensors to artifacts, and every stored asset link re-signed per request so URLs survive storage-endpoint migration without a data migration; executed 10,700+ inference jobs across 2,500+ completed multi-stage runs, generating 3,300 conservation risks and 3,900 LLM insights.
  • Built the telemetry layer on TimescaleDB hypertables (composite time+UUID keys) with a GraphQL API unifying historical/forecast/anomaly reads across sensor, zone, floor and building scopes, sustaining 3.6M rows over 119 days of continuous ingest and serving 50K-row, 30-day analytical windows in ~250 ms, plus a fleet-health engine deriving battery depletion (regr_slope 30-day projection), persistent weak links (p10 RSSI, not min) and per-series dropout gaps in raw SQL, surfacing failing sensors the static config columns reported as 97% healthy.

Delivery1

  • Owned the delivery stack: 5-process Docker Compose topology (ASGI + beat + three queue-specialised workers), Jenkins CI, a tmux dev orchestrator, and an idempotent setup script that health-checks Postgres/TimescaleDB/Redis/RabbitMQ/S3 and provisions vhosts and quorum queues from the application's own queue definitions, so broker topology cannot drift from code.