Team
dronomy.io
Project Concept
AdaBoost AI — Hybrid Hawkes Router
Score: 85.6% | F1=1.00 | 100% on-device | ~600ms avg
FunctionGemma 270M runs every query (hackathon requirement)
NLP validates/rescues FunctionGemma’s broken string args
Hawkes process (λ(t) = μ + Σα·e^(-β·Δt)) governs routing decisions
Cloud fallback (Gemini Flash) for truly unknown patterns (threshold > 0.15)
Key files: src/agents.py, src/nlp_extract.py, src/router.py
Entry
Status: Submitted
Last saved: February 21 at 5:24 PM PST
Team Roster
You must be registered for the event to view the team message board.
Suvasis Mukherjee Team Lead RSVP Approved
owner at adaboost
1 person team. I was responsible for all aspects of the project as a solo participant:
HYBRID ROUTING ARCHITECTURE:
Designed and implemented the Hawkes process-based routing algorithm (λ(t) = μ + Σα·exp(-β·Δt)) — a self-exciting point process borrowed from high-frequency trading — to dynamically govern edge vs. cloud routing decisions. The router calculates per-query complexity scores and uses the Hawkes intensity function to determine whether to route to on-device FunctionGemma or cloud Gemini Flash.
SPONSOR TOOLS — CACTUS SDK & FUNCTIONGEMMA (Google DeepMind):
Integrated FunctionGemma 270M via the Cactus Python SDK (cactus_init, cactus_complete, cactus_reset) as the primary on-device inference engine. FunctionGemma runs on every query for tool-call generation. Discovered and worked around FunctionGemma's nondeterministic behavior on string arguments (~67% success rate) through systematic debugging, including using cactus_reset() between queries to prevent KV cache degradation that caused latency to balloon from 300ms to 2800ms. Tuned max_tokens and force_tools parameters for optimal speed.
SPONSOR TOOL — GEMINI 2.5 FLASH (Google DeepMind):
Integrated Gemini 2.5 Flash as the cloud fallback for complex multi-tool queries (3+ function calls). Used Gemini's function calling API with full tool schemas. The cloud path activates when query complexity exceeds the Hawkes-governed threshold, providing a genuine hybrid edge-cloud architecture.
NLP VALIDATION LAYER:
Built a general-purpose regex-based NLP extraction module (src/nlp_extract.py) that validates, augments, and rescues FunctionGemma's output. The NLP layer handles 7 tool types (send_message, search_contacts, create_reminder, get_weather, set_alarm, set_timer, play_music) with broad pattern matching that generalizes beyond exact benchmark phrasings. This addresses FunctionGemma 270M's fundamental limitation with string argument extraction (full-width colons, escape tags, corrupted values).
KEY TECHNICAL DECISIONS:
- FunctionGemma always runs first (satisfies hackathon requirement for on-device model usage)
- NLP compares its extraction against FunctionGemma's output: if NLP finds more calls (multi-tool queries FunctionGemma missed) or same count (more reliable for string args), NLP result is used
- Hawkes process parameters (μ=0.15, α=0.25, β=1.5) tuned for optimal edge-first routing
- Complexity threshold (0.15) set to maximize on-device processing while keeping cloud path active for truly unknown patterns in held-out evaluation
RESULTS: 85.6% score — F1=1.00 (perfect accuracy), 100% on-device, ~600ms average latency across all 30 benchmark queries.
Suvasis Mukherjee is the Owner at adaboost (dronomy.io) and is self-employed. With over 20 years of experience, Suvasis is seeking investors for fundraising and is open to introductions via email. Their background includes education from Stanford University, Indian Institute of Technology, Roorkee, and MSCS from Santa Clara University and graduate certification in fields like Artificial Intelligence from Stanford Univ. Suvasis is also a Tinkerer Advisor/Contributor/Angel (Technical) who built MEV Shield to detect hidden liquidity cliffs in Uniswap V3 pools.
DeFi, MEV Shield, Uniswap V3 liquidity analysis, autonomous robotics, edge AI, distributed systems, NLP, computer vision, deep learning.
MEV Shield detects hidden liquidity cliffs in Uniswap V3 pools before you trade.
When LPs remove liquidity at specific price points, slippage suddenly spikes —
MEV bots exploit this to sandwich your trades. We analyze tick-level topology
from Mint/Burn events, identify where liquidity drops 50-90%, and warn you:
"Your $50K swap crosses 3 cliffs — expect 5× worse slippage."
Built with 20 years of Wall Street microstructure expertise.
See the cliffs before the bots do.