backend & AI engineer, Gurgaon
Divyanshu
Aryan
I live in Gurgaon. I love movies, play games, never miss a Formula 1 race, and listen to Fred again.., Travis Scott and Drake a little too loud. These days I build AI that people in India talk to, at TrulyMadly.
01 — where I live
Gurgaon, as I know it.
Glass towers, endless traffic, and a few spots that feel like mine.
Glass and dust
Half the city is shiny towers, the other half is still being built. Both are on the way to work.
Galleria & Effingut
My favourite corner of the city, close to home and close to the office. At Effingut, order the mango cider.
Weekends are for Delhi
Gurgaon gets the weekdays. Come the weekend, I’m usually somewhere in Delhi.
Burgers, kebabs, chicken
Burgers from Aku’s, kebabs, and anything with chicken in it.
02 — movies
I love movies. Probably too much.
The films and series that stayed with me.
- Interstellar
- The Dark Knight trilogy
- Hachi: A Dog’s Tale
- Dark
- Game of Thrones
- Breaking Bad
- favourite director
- Christopher Nolan
- the ending I hate
- Game of Thrones, season 8
03 — games
And when I’m not watching, I’m playing.
most hours played
- The Witcher 3: Wild Hunt171 h
- Grand Theft Auto V112 h
- Cyberpunk 2077102 h
play this once
Life is Strange
04 — formula 1
Sundays are for F1.
Race weekends are sacred: lights out, the first corner, the strategy calls, the last-lap chaos. And yes, I’m a Lewis Hamilton fan.
the only number that matters
“Still we rise.”
05 — music
Music is how I keep time.
This part is live, straight from my Spotify.
All-time favourites: Fred again.., Travis Scott and Drake. Also always on: Cigarettes After Sex, Gregory Alan Isakov and Seedhe Maut.
on repeat this month
albums that made me
Between FlowersOAFF
Euphoria (Original Score from the HBO Series)Labrinth
Modern PunjabHari & Sukhmani
USBFred again..
Why I build what I build
That’s the life part: movies, games, F1 and a lot of music.
Now the work. At TrulyMadly, an Indian dating app, I build the AI people talk to, and the systems that decide who they meet.
06 — how I got here
From circuits to AI.
- 2021
Birla Institute of Technology, Mesra
Electronics & Communication, in the hills outside Ranchi.
- in college
Playing with early LLMs
Building small things with the first models I could get my hands on. That’s when code took over from circuits.
- Jan 2025
Intern at TrulyMadly
Built an internal chat platform where staff could switch between AI models mid-conversation.
- Jul 2025 → now
Backend & AI engineer at TrulyMadly
Recommendation scoring, then Maira, TrulyMadly’s AI matchmaker, as its sole engineer.
07 — at TrulyMadly · ranking & ML
I built the scoring behind your matches.
Everything on a dating app starts with whose profile you see next. At TrulyMadly I built several of the scoring systems that make that call, proving each idea as a small prototype before it went live. One, which I built end to end, predicts which new users will subscribe. Together they improved the matches and grew subscriptions.
likelier than average to subscribe: the top 1% of new users my model picks
for engineers
- End-to-end model predicting new-user paid conversion: dataset, feature engineering, XGBoost (0.78 ROC-AUC, 7× top-1% lift), deployed into production recommendations.
- Multiple scoring services deployed in the recommendation engine, improving ranking quality and subscription growth; POCs to validate features before rollout.
- Event-driven enrichment pipeline turning profiles into vision-LLM attributes and persona embeddings; migrated vector search from MongoDB Atlas to Qdrant behind an embedding microservice.
08 — at TrulyMadly · AI agents
I built Maira, TrulyMadly’s AI matchmaker.
People rarely say what they want neatly, and many say it in Hinglish. As the sole engineer, I designed and built Maira (it started out as MissTM) to follow that, learn what each person wants and introduce real people who fit, with safety checks throughout. Then I rebuilt how it reads requests: it now understands more, works faster and costs far less.
koi chahiye jo travel karna pasand kare, aur thoda funny ho
MairaSomeone who loves a trip and doesn’t take life too seriously. Got it. Delhi NCR mein dekhun?
requests it understands correctly, up from 77 before my rebuild
of the old cost to understand each request
less waiting while it works out what you mean
for engineers
- Sole engineer. Agent harness: layered prompt assembly (persona, constitution, time awareness, conversation arc), intent routing, output guards, 6 LLMs with failover and A/B tests.
- Rebuilt Hinglish preference extraction as one cached LLM tool call: accuracy 77% → 96%, cost −82%, latency −37%.
- Search: hard filters, hybrid retrieval over text and image embeddings fused with RRF, a reciprocal likes-you channel, two-sided reranking with exposure fairness.
09 — at TrulyMadly · memory systems
Teaching an AI to remember you.
Maira grew into an app of its own, where people chat with and call AI characters, and I architected it. The hard part was memory, so I designed a layered one and measured it on real chats. Tell a character your sister’s wedding is in March, and weeks later it asks how it went.
3 weeks agomy sister’s wedding is in march, total chaos at home
todayHow did the wedding go? Did the chaos settle down?
personal details it picks up from real chats, before → after my redesign
for engineers
- Architected the standalone Maira companion app (chat and calls with many human-like characters) on a multi-tenant platform that drove revenue growth.
- 3-tier memory: recent window, compaction summaries, RAG over subject–relation–object triples extracted async into a knowledge bank. Prompt caching serves ~84% of tokens.
- LLM-judge evaluation on real chats; designed an extractor lifting user-fact recall from 19% to 74%.
10 — at TrulyMadly · re-engagement
I designed check-ins people actually answer.
When someone goes quiet, the easy answer is spam. I designed something gentler: the character checks in with something it remembers about them, waits longer each time, from two minutes up to fifteen days, and tries a fresh angle on every attempt. Then I rebuilt it to hold up at scale, so every message goes out once, and only once.
people who got a check-in wrote back
for engineers
- Re-engagement pipeline: memory-aware nudges on a back-off ladder (2 min → 15 days), 19 rotating comeback angles, plus story and feed nudges; 70%+ reply rate.
- Re-architected for scale: lease-based dispatcher, transactional outbox, idempotency fences. Exactly-once delivery at 2.5 s p95.
11 — at TrulyMadly · real-time voice
Making an AI you can actually call.
Typing isn’t for everyone, so the characters take calls too. A call only feels real if replies come quickly and you can cut in. I connected a live AI voice model to the app’s calling system, made it stop the moment you start talking, and carried the chat memory into the call.
for engineers
- Real-time voice calls on Gemini Live through a Go bridge to Agora, sharing memory with the character’s text chat.
- Barge-in with Silero VAD, so people can interrupt mid-sentence, as on a normal phone call.
12 — at TrulyMadly · platform & automation
The quieter systems I built.
Not everything I build is a feature you see. Some of it makes the app cheaper to run and more dependable.
A seventh of the cost per support ticket
I built an AI support agent that sorts tickets, finds the account, checks payment screenshots and drafts or sends the reply, keeping personal data protected. Replies come three times faster.
Check-ins sent twice
When I rebuilt check-ins for scale, I made every message go out exactly once. 95 in 100 leave within 2.5 seconds.
for engineers
- Support automation: helpdesk triage, account resolution, vision-LLM payment-proof checks, drafted or auto-sent replies with PII guardrails. An 18-tool MCP service drives a tool-calling agent: 7.5× cheaper per ticket, 3× faster.
- Async event-driven services on BullMQ workers and signed webhooks, so APIs never block on LLM calls; per-session and per-feature LLM cost tracking with internal React dashboards.
- 7+ microservices on Docker and Kubernetes via GitLab CI/CD and Argo CD; BigQuery ETL, Sentry, New Relic. Also AI icebreakers, voice-note transcription, A/B-tested image moderation.