I build AI agents that have to survive production, and I publish the measurements.
Isolated execution on Firecracker microVMs. Decision models that answer in 100 ms with calibrated confidence. Governed memory. Robots that take orders in plain English. Everything below links to a repo, a benchmark run, or a merged PR.
|
Agentic AI on Bedrock and AgentCore, migrations to Bedrock (OpenAI, Azure OpenAI, Vertex), retrieval over large corpora, Amazon Q / Quick. Architect who ships the code and the benchmark. |
Open an issue first, then a PR. Most useful right now: microvm-ctl (runtimes, fleet scenarios) decider-lab (adapters, eval sets) memwarden (poisoning detectors) |
Talks, workshops, AWS community events, or a question about anything here. vivekraja98 [at] gmail [dot] com Bengaluru / Madurai, India |
| 3.54 s | 111 ms | 93.8% | 57,796/s | 0.7 s | +23.4 |
|---|---|---|---|---|---|
| microVM launch to serving traffic (p50) | warm request inside the VM (p50) | cheaper than always-on for a 30 min + 8 h session | governed memory writes per second, 0 poisoned records admitted | first request wakes a suspended VM | Intelligence gain from a 150-step decider fine-tune, 95% CI [18.2, 28.7] |
Every number is from a recorded run in the linked repos, not a quote from a datasheet.
Each agent gets its own Firecracker VM and kernel. Strands Box runs inside and checks every shell command, file touch and outbound request against a policy. I wrote the control plane, the fleet runtime, the playground and the benchmarks.
|
microvm-ctl · PyPI · Apache-2.0 Build a snapshot image from a Dockerfile, launch Firecracker microVMs, scale fleets at 80% of applied quota, call each VM over an authenticated endpoint, hand tasks over via a lease from Step Functions or Lambda durable functions.
|
benchmarks/benchmark.py.
|
||||||||||||||
|
|
strands-box-microvm-ctl · live playground Strands Box built for aarch64 Linux, one box per task, fleet scenarios, and a playground behind CloudFront + WAF. Also contributed Linux support upstream to Strands Box.
|
||||||||||||||
|
microvm-handoff-demo Nine scenarios run live on both Step Functions and Lambda durable functions, with execution histories, VM logs and console screenshots checked in. An 8-shard fan-out completes in 8.9 s for $0.00031 per run. The first pass found two bugs that became microvm-ctl 0.3.1. awesome-microvm 14 runnable examples (code sandbox, agent eval fan-out, notebook sessions, DuckDB over S3, CI runner, multi-tenant agents, circuit breaker) and a 5-part article series on AWS Builder Center. |
|
Articles: Control and scale Lambda MicroVMs with microvm-ctl · Seven workloads Lambda could never run until MicroVMs · A kernel for every customer: 1,000 tenants · Hand a task to a MicroVM from anywhere
Strands Decider is a 2B "system one" model: one forward pass, no decoding, a calibrated probability on every yes/no, choice or score question, ~115 ms on a 3090. I work on making it see, making it better, and making it deployable.
|
Giving the decider eyes, with zero image training
Qwen3.5-2B-Base ships a vision tower nobody was using. I remapped v19's LoRA onto the multimodal torso and put images inside
Matches an image-trained 2B on accuracy (paired bootstrap CIs cross zero), with 5x lower calibration error. Cost of the whole eval: one H100 for 35 minutes, $1.58. Then training it on 38k image rows (VQAv2 pairs, COCO counting, Mind2Web, own chart and invoice renders), 3 seeds, ViT frozen, ~56 min per run:
Better on every image benchmark and every seed, no regression on 899 held-out text rows (0.817 -> 0.829). The blind-image column matters: before training, the model stayed confident with the image removed. Now it hedges. |
Training better text deciders Base bake-off at 2,400 steps: MiniCPM5-2B beats Qwen3.5-2B by +13.0 JevBench tasks (CI +6.3, +20.0). My v21 candidate on MiniCPM5 scores 179/231 on the public set vs the published v19's 167, a paired gain of +12.3 tasks (CI +3.0, +22.3). The honest part: on 1,050 rows from five families no decider had seen, MiniCPM5's lead evaporates (-17 to -21) and a Qwen weight soup wins. Every model I trained still beats the two competing open 2B deciders on those unseen families (+2.6 to +14.5). All of this is preregistered, three seeds per arm, and labelled exploratory until a confirmatory run. Full results |
|
|
|
Upstream to strands-labs/robots (150+ robots, 8 categories): 9 merged PRs authored, 3 co-authored. The ones I care about most: rosbridge transport with RosbridgeRobot.from_curiosity() (#1110), ROS 2 action support and goal-level navigate_to (#995), AWS DeepRacer as an Ackermann ROS 2 robot (#2689), FastTD3 as a native off-policy RL provider (#3206).
memwarden wraps any agent memory backend (AgentCore Memory, Redis, pgvector) with tenant isolation, per-category retention, provenance-based trust tiers, two detection layers, four fail-closed read gates, a hash-chained audit log and GDPR Article 17 erasure certificates. Targets OWASP ASI06. Companion to the Engram paper.
Lab run, 6 tenants, 3,181 writes, 23% poisoned: 57,796 governed writes/s, 0.017 ms p50 overhead, 550 blocked at write, 188 quarantined, 0 adversarial records reached agent context. The README says plainly that the inline detector is a precision instrument (held-out paraphrase recall 0/10) and the trust gate carries the guarantee (7 of 8 held-out attacks stopped at read).
|
|
f1-agentic-analysis · live
Deterministic pandas/SciPy computes everything (Theil-Sen degradation fits, pit loss, undercut ledger). Seven Strands agents with assay · live Read-quality scoring for AWS Builder Center. Over 302 articles in one week, likes vs quality correlated at r = -0.09, and the top 20 by likes and top 20 by score shared zero articles. 96% exact verdict agreement with a hand-labelled golden set. aws-samples / OpenClaw on AWS with Bedrock · ★ 415 Contributed the AgentCore Runtime deployment path (PR #16), since superseded by the maintainers' multi-tenant build. Presented at ClawCon Seattle 2026. Also clawskills-strands: 52 OpenClaw skills converted to executable Strands tools. |
Papers. Engram (memory governance, with memwarden), One Agent Many Bodies and Design Principles of Agent-Powered Robotics Systems (robotics, simulation-validated, arXiv pending). 7 peer-reviewed papers, 3 granted patents, across AI and brain-computer interfaces.
Before Aivar. Five years as VP of Products at Nexstem, a brain-computer interface startup: EEG hardware with motorized electrodes, firmware, signal pipelines, on-device inference, SDK. The path ran silicon -> firmware -> embedded -> cloud -> ML -> agents -> robots, which is why I care about what sits under the model.
Community. Founded AWS User Group Madurai; help run communities across Madurai, Chennai, Coimbatore and Bengaluru. 80+ talks including AWS re:Invent, 500+ meetups. AWS Machine Learning Hero at 22, now AWS AI Hero.
Bedrock AgentCore Strands Agents Strands Decider Strands Box Firecracker Lambda MicroVMs ROS 2 Nav2 OpenSearch CDK Python TypeScript
I occasionally build things because somebody tells me they probably shouldn't work.