Software Engineer · Backend, Systems, AI · Gurugram, India
Saurabh
Singh
“I build long-lasting systems for fun, and I know when to ask the right questions.” Founding engineer at Temple. Previously Zomato, MPL, Signzy.
SCROLL ↓ NOW · TEMPLEFounding Engineer · Sept 2025 – present
The systems the product runs on.
A health company building a precision wearable worn at the temple. The placement gives cleaner physiological signals, which the platform turns into sleep and long-term health insight.
- BLE parser
- The wearable streams sensor data over Bluetooth Low Energy as compact binary packets. This layer decodes them into typed sensor streams, handling framing, sequence ordering, and malformed payloads, so nothing downstream ever touches raw bytes.
- The device path
- The pipeline that moves device data from ingestion to storage: readings land on a Kafka backbone and settle into a per-user time-series store. Real-time and bulk-sync traffic share one event-driven spine, so a live stream and a week of backfill follow the same path. The work spans firmware and backend.
- CDC into Kafka
- Change data capture from the primary datastore into Kafka: row-level changes published as ordered events. Downstream services react to data as it changes instead of polling, and the primary store stays isolated from read load.
- The Sleep system
- Assembles a night of sleep from raw signal streams. Data arrives late and out of order due to delayed syncs and connectivity gaps, so the system merges idempotently, infers sleep and wake boundaries, and reconciles conflicts down to exactly one correct record per user per night.
- Timezone model
- A platform-wide rebuild of time handling. Every record is resolved to the wearer’s local timezone at one well-defined point, with a single model shared by all services. That closes the class of bugs around travel, DST shifts, and data that crosses midnight.
- Insights
- Asynchronous pipelines that turn accumulated health data into daily insights and check-ins. Statistical layers establish personal baselines and flag deviations; LLM stages turn them into readable summaries. Everything runs off the request path, so generation never blocks the product.
- Cloud-to-wearable
- A signed command channel from cloud to device over Bluetooth, using SMP as the transport protocol and CBOR as the wire encoding. Every command is ECDSA-signed against a per-boot nonce, so it cannot be forged or replayed. The design spans firmware, both mobile apps, and backend.
- Firefighting bots
- Automated monitors across every service in the fleet, watching health, error rates, and data-flow anomalies. Issues are raised with debugging context attached, before users are affected.
- Remote debugging
- A setup that attaches live breakpoints to services running in the cloud, directly from a laptop. Production issues get inspected where they happen, with real state and real traffic, instead of being reconstructed locally from logs.
- Internal tooling
- Purpose-built internal tools for debugging, data inspection, and routine operations: inspecting device data safely, tracing a record through the pipeline, verifying a fix in production.
STACK Go · TypeScript · Python · C · Kafka · gRPC · BLE/SMP · Docker · AWS · GCP
Software Engineer · May 2024 – Sept 2025
AI infrastructure for 500+ engineers.
India’s largest food-delivery and dining platform, and a publicly listed company. Think DoorDash or Deliveroo, at India scale.
- AI Gateway
- A single gateway in front of six AI providers: routing and load balancing across them, per-client rate limits, retries with fallbacks. Teams integrate once and stay insulated from individual providers’ quirks and outages.
- Prompt platform
- 1000+ prompts across 100+ projects, versioned and served at runtime instead of hard-coded, so a prompt change ships without a code release. Releases got 50% lighter, and teams iterate on AI behavior independently of the deploy cycle.
- Streaming
- Real-time streaming of LLM responses: tokens render as they are generated rather than after full completion, with connection lifecycle handled at the platform layer. Holds 200+ concurrent sessions.
- Cost tracking
- Automated attribution of AI spend across 4 tenants and 6 providers, replacing manual reporting entirely. Alongside it, an assertions service that checks model outputs, attaches live citations, and applies moderation before responses reach users.
0K+ requests a minute through one gateway 0% uptime at ≤100 ms −0% latency after my streaming rebuild
STACK Go · React · Python · Docker · AWS · Prometheus · GCP
2020 – 2024 · two internships and a B.Tech in Computer Science
Two internships, one degree.
A live game, as an intern.
- Draw 4
- A live multiplayer card game with players across countries. Backend feature work on its Node.js services, including moving inter-service calls from REST to gRPC, which cut response times.
- Job schedulers
- BullMQ over Redis for game workflows: delayed jobs, retries, and recurring tasks that must fire reliably, built so a retried job cannot double-fire.
- Observability
- Kibana over structured logs: tracing issues across service boundaries during incidents.
Video KYC, without a human.
- Unassisted VKYC
- The flow banks run when someone opens an account remotely: capture over live video, automated checks, and a decision, with no human on the call. Backend work on the Node.js services orchestrating that session.
- Release automation
- Playwright end-to-end suites on every build; a failure blocks the release, so regressions surface before customers see them.
B.Tech, Computer Science · IIIT Una ↗ · Dec 2020 – May 2024 · one of India’s national Institutes of Information Technology, entered through JEE, an exam taken by over a million students a year
CGPA 0 / 10 · JEE Mains 0%ile · Flipkart Grid 4.0 · Publicis Sapient Jumpstart · 4 sports organized on campus
STACK Node.js · TypeScript · Redis · Playwright · Kibana · C/C++ · SQL · AWS · Docker
Let’s talk.
Built this myself. Everything above shipped.
Gurugram, India. Open to work anywhere.
Logos belong to their respective companies · plain version