KD

The Kalyanam Archive · Vol. I

Some things should be built quickly.

The important ones should be built properly.

Kalyanam Dewri

Guwahati-born. ASU-forged. Built properly.

Enter the ArchiveFor recruiters →

↓ Or scroll

Live · four worlds, one operator

Trading

8.7M msgs/sec · 87µs

Autonomy

V2X consistency

AI

Grounded · 83→95%

Founder

Steeped in heritage

Live proof170Mrecords / day·8.7Mmsgs / sec·87µsp99 latency·28knightly tests·83→95%grounded resolution·38%lower inference $·

01 · Receipts

Proof, when you want it.

The archive keeps receipts: systems shipped, tests run, traces measured, demos built, and public proof where it exists. The full set lives in the dossier and case studies. These three are the ones I would defend in a room.

Matching engine
8.7M
msgs / sec

Canyon Exchange C++ order matching.

Data infrastructure
170M
records / day

Kafka / Flink quality dashboards at Blue Yonder.

Autonomy regressions
28k
nightly tests

BELIV Lab replay / simulation harness.

02 · Identity

Three disciplines.
Same hand.

Most engineers specialize in one. These three have all shipped, under the same hand. The rooms next door are open if you want to wander.

01 / 03

AI Applications

Recruiter & candidate assistants, RAG, tool-calling, evaluation harnesses, latency / cost optimization at production scale.

11k+ weekly assisted interactions
02 / 03

Systems & Infra

Go / TypeScript APIs, Spark / Arrow, Kafka / Flink, C++ matching engines, eBPF and perf-driven hot-path tuning.

170M records / day · 8.7M msgs / sec
03 / 03

Autonomy & Robotics

ROS2, C++, PyTorch, CUDA, replay simulation, V2X sensor consistency, regression harnesses for connected-vehicle stacks.

97 → 11 min reproduction

03 · Selected work

Selected work.
The doors are unlocked.

C++ · Low-latency systems · Linux

Canyon Exchange

A from-scratch matching engine that moves 8.7M messages per second at 87µs p99 — the crown jewel for systems / trading roles.

8.7M msgs/sec87µs p99Lock-free + io_uring

ROS2 · CUDA · V2X · Connected vehicles

RoadTwin

Replay simulator and consistency models for camera, lidar, and V2X streams under packet loss and sensor desync.

Small-object recall ↑Replay regression setNight / desync modes

ROS2 · C++ · Connected-vehicle perception

BELIV Replay Harness

Cut nightly scenario reproduction across 28k tests from 97 minutes to 11 minutes for ASU's connected-vehicle research lab.

97 → 11 min28k nightly testsV2X consistency models

RAG · Tool-calling · Evals · Production AI

AI Assistant Platform

Eval-driven recruiter / candidate assistant flows: grounded resolution 83% → 95%, latency −34%, inference spend −38%.

11k+ weekly chats83 → 95% grounded−38% inference $

Spark · Arrow · Kafka · Flink · 140 TB / mo

Blue Yonder Data Observatory

Throughput +2.1×, compute cost −27% across 140 TB/month transformation pipelines; SLO dashboards over 170M records / day.

140 TB / mo+2.1× throughput41 → 12 min triage

AI on CCTV + radar · Devils Invent / DASSH winner

Live Emergency Response System

Top $10K prize at the 2023 Devils Invent + DASSH hackathon: AI in CCTV streams and radar mapping for first responders.

$10K top prize30-hour buildCCTV + radar fusion

04 · Experience

2026CitadelSoftware engineering — systems
2025ParadoxAI product engineering
2024GoDaddyPlatform engineering
2024Blue YonderData infrastructure
2024–2025BELIV Lab @ ASUAutonomy research engineering
2025RoadTwinFounding engineer — perception & V2X
2025–2026Canyon ExchangeMatching engine — C++ / Linux
2024–Assam Ice Tea by Dewri'sFounder

05 · Ask

The archive answers back.
Ask it what survived.

Open the archive →Or just email me

Vol. I · 2020 – 2026. Vol. II forthcoming.
The unreleased material remains sealed.