00 - vladislav kondratyev
dallas, texas · cdt
ai systems engineer · inference · runtimes · agent tooling
ut dallas cs (class of '27). i build ai runtimes, inference pipelines, and developer tooling across vr, graphics, and cloud systems. recent work spans meta reality labs, intel xess, nutanix, and md7.
01 - about
i'm a cs student at ut dallas (class of '27) focused on ai systems work that has to perform outside a notebook. that has meant real-time inference in vr, model deployment and optimization around intel xess, and backend systems that stay responsive under production load.
the throughline is systems thinking: model optimization, async services, low-latency data paths, and tooling for agentic workflows. outside internships and contracts, i publish open-source tooling, build products, and have picked up 11 wins across 18 hackathons. check out the projects if you want to see the work directly.
languages
ai / ml
systems
devops / cloud
02 - experience
VR & AI Engineer · austin, tx · jan 2026 → apr 2026
embedded inside meta's reality labs to architect and ship a real-time ai-powered vr system within meta horizon os. integrated pytorch inference models into immersive vr environments achieving <100ms response time. built multimodal ai pipelines fusing user input, environmental context, and behavioral telemetry into a shared transformer-based embedding space.
<100ms
vr response latency
Software Engineer Intern · austin, tx · may–aug 2025
worked on xess 2.x, intel's flagship ai super-resolution technology shipping inside aaa game pipelines. developed ai-powered super-resolution shaders in hlsl and dpc++, reducing temporal artifacts by 25%. refactored model deployment pipeline (c++, pytorch, openvino, cuda), reducing inference latency by 35%. exported pytorch models to onnx and tensorrt with int8 quantization - 40% size reduction with <2% quality loss.
40%
faster frame rates
Software Engineer II · san antonio, tx · aug 2024 → may 2026
promoted within 11 months. designed distributed real-time decision pipelines using async event-driven architecture (python, java, kafka), improving throughput by 20% and reducing p99 latency by 35%. engineered low-latency data layers handling 15,000+ concurrent users. diagnosed linux-based production bottlenecks via perf/flamegraphs, reducing downtime by 25%.
15,000+
concurrent users
Founding Software Engineer · bronxville, ny · sep 2024 → apr 2025
co-founded an early-stage fintech startup that was successfully acquired. built a scalable saas financial tracking platform from zero serving 1,000+ users and processing $300,000+/month. cut latency by 40% using express.js and mongodb. achieved 99.9% system availability with github actions + docker ci/cd pipelines.
$300K+
monthly transactions
AI Systems Engineer Intern · remote · jan 2026 → present
working on ai systems infrastructure at nutanix with a focus on production integration, reliability, and enterprise-scale engineering constraints.
enterprise
ai infrastructure
03 - by the numbers
03 - hackathons
11x winner across 22 competitions. 5 wins in 3 weeks. hover each shard to see details.
HackHCC · may 2026
BrainDance
eeg-based memory reconstruction, viewable in vr.
winnerclick to open post →
LA Hacks '26 · apr 2026
SAGE
ai education platform: real-time physics, on-device inference, live learner presence.
cloudinary track winnerclick to open post →
Hook 'Em Hacks '26 · apr 2026
ClawGuardian
decentralized prompt injection firewall for ai agents with on-chain threat sharing.
2x track winnerclick to open post →
HackSMU VII · apr 2026
Project Tusk
real-time bioacoustic pipeline for elephant field recordings.
winnerclick to open post →
Global Hack Week · mar 2026
CAFAI
context-aware fused ad insertion: ads that blend into content instead of interrupting it.
winnerclick to open post →
hover a shard
5 of 11 wins shown, each shard is a real hackathon win
selected system
invented adaptive pattern-exploiting sorting algorithm.
most sorting algorithms assume nothing about the input. sherlock-sort opens by profiling the data - checking for near-sortedness, clustering, repetition, and distribution shape - then selects and combines proven algorithms (insertion, merge, radix, counting) in the optimal mix for that specific input.
selected system
multi-llm council orchestration with cli and fastapi.
any single llm can confidently return a wrong answer. ai-decision-council routes every query through multiple models simultaneously, collects their responses, and runs a configurable consensus algorithm before surfacing a result - dramatically reducing hallucination rate on factual and reasoning tasks.
packaged and published to the python package index so anyone can install it with a single pip command. includes full cli with interactive mode and a programmatic python api for embedding into larger pipelines.
selected system
prompt-injection firewall for ai agents with on-chain threat sharing.
as ai agents gain tool access and autonomous action, prompt injection - tricking an agent into ignoring its instructions via malicious input - becomes a critical security vulnerability. clawguardian intercepts every input before it reaches the model and classifies it against a continuously updated threat database.
when clawguardian detects and blocks an attack, it logs a hashed threat signature to a shared on-chain registry. every other clawguardian instance pulls this feed and updates its local classifier - creating a network effect where novel attacks discovered by one node are immediately defended against by all.
selected system
cognitive ai system built around persistent memory, emotional modeling, and real-time context ingestion.
echomind fuses audio, vision, and system telemetry in real time. each stream is encoded independently then merged into a shared embedding via attention-weighted cross-modal fusion. the result is an agent that doesn't need to be told what you're doing: it perceives it directly, updates its context window, and adapts its behavior without full recomputation.
most ai companions reset on every conversation. echomind maintains a sliding-window episodic memory buffer with salience-weighted retention: informed by the sect model research paper authored under echolab. the agent remembers what matters and forgets what doesn't, the way humans do.
selected system
elephant vocalization noise-removal and research platform.
elephant infrasound communication is critical to conservation research, but field recordings are dominated by wind, rain, and human noise. project tusk applies targeted dsp pipelines tuned to the 14–35 hz infrasound range to strip noise without degrading the signal researchers care about.
selected as a winning project at hacksmu '26 for demonstrating a practical application of signal processing to a real conservation problem. the platform was built end-to-end in 24 hours by a small team.
selected system
local runtime and control layer for autonomous ai agents.
most agent frameworks require a hosted service. gagent runs entirely on localhost - persistent task queues, registered tools, and a memory layer that survives restarts. agents can call any local or remote tool you register and their state is checkpointed so long-running tasks survive crashes.
05 - wins
10 competitions. 11 wins total. intel and nutanix wins under nda.
Internal innovation challenge. Details under NDA.
Internal engineering challenge. Details under NDA.
Context-aware fused ad insertion: ads that blend into content instead of interrupting it.
AI education platform with real-time physics simulations, on-device inference, and live learner presence.
Real-time bioacoustic pipeline for elephant field recordings: noise removal, feature extraction, AI classification.
Prompt-injection firewall for AI agents with on-chain threat sharing. 2x track winner: IBM Security + AWS 1st place.
EEG-based memory reconstruction pipeline viewable in VR.
AI-powered lecture-to-study-resource assistant transforming recorded lectures into structured study materials.
Hackathon win at UTSA's RowdyHacks competition.
06 - research
ut southwestern medical center · ut dallas
working with Dr. Fang at UT Southwestern Medical Center, I built machine learning pipelines that distinguish malignant cells from healthy tissue in high-resolution imaging data, the kind of classification that normally requires a trained pathologist and hours of manual review.
the system extracts morphological features at the sub-cellular level, classifies cell populations across imaging slices, and flags anomalies consistent with early-stage malignancy, turning raw microscopy output into structured, actionable signal for researchers.
what it means
early detection is the single largest factor in cancer survival outcomes. automating the visual triage of tissue samples compresses the timeline between imaging and diagnosis.
07 - github
08 - open source packages
gorchestrator
Parallel agent execution manager: runs N attempts, scores outputs, selects the winner
-
last 30d
gmirror
Synthetic-user verification layer: scores LLM outputs against cognitive user panels
-
last 30d
Meta - Reality Labs
CNF Technologies