Állysson Oliveira
Tech Lead — Generative & Agentic AI Engineer
Transforming frontier foundation models into hardened enterprise reality. Specializing in IBM Spyre AIU hardware acceleration, the CARROT Pareto-optimal LLM routing engine, and foundational data pipelines that power IBM Granite. Proven track record safeguarding multi-million dollar contracts and securing on-time releases for mission-critical System Z mainframes.
Chat with Állysson's Digital Twin
An autonomous conversational agent grounded in Állysson's verified career history, research publications, patents, and silicon co-design at IBM Research.
Állysson's Digital Twin
Qwen 27BOnline • Grounded in Profile.pdf & IBM Research
Signature Strategic Impact
Real enterprise leverage: safeguarding revenue, optimizing agentic AI on proprietary silicon, and moving millions in mission-critical contracts.
Rescuing WXA4z on Spyre AIU & Securing ~70% of System Z Sales
When strict AIU hardware constraints threatened to delay the release of WXA4z on Spyre, Állysson engineered prompt optimizations and model adaptations that catapulted Granite-based agents from sub-20% accuracy to ~90% of frontier Llama performance within weeks. He repeated this breakthrough on IBM Power, boosting accuracy from 50% to 90% to enable the first Power GA with Spyre and paving the runway for OpenShift AI on Power.
Saving Verizon Watsonx.ai Contract & Unlocking ~$3M New Revenue
Facing contract cancellation on Watsonx.ai due to cross-border legal compliance constraints surrounding third-party models, Állysson rapidly benchmarked, fine-tuned, and optimized Granite's performance by over 50%. The result met all regulatory mandates and technical SLA thresholds, securing Verizon's long-term commitment and generating ~$3M in new revenue.
CARROT: Pareto-Optimal LLM Router & 1M+ Evaluation Engine
Co-authored the CARROT research paper and engineered a massive messaging queue platform evaluating models across Open LLM Leaderboard datasets, commercialized directly in Watsonx.ai.
Consolidating IBM Consulting Advantage onto Granite 3 8B
Migrated fragmented enterprise assistant workflows from expensive third-party models onto IBM Granite 3 8B across the flagship IBM Consulting Advantage (ICA) platform.
IBM Data Pile: Engineering the Pretraining Ingestion Engine
Built the multiprocessing web-scale crawler and data gathering pipelines for IBM Data Pile (formerly Blue Pile), the core dataset repository used to pretrain IBM Granite.
Career Journey & Evolution
From low-level C/C++ telecom protocol stacks and mobile operating systems to leading hardware-accelerated Agentic AI and foundation model research at IBM.
- Prevented a delayed release of WXA4z agents on Spyre (IBM's custom AIU), protecting revenue tied to ~70% of new System Z mainframe sales.
- Solved major model constraints based on Spyre requirements by improving Granite-based agents from sub-20% accuracy to ~90% of Llama performance within weeks, enabling a viable on-time release.
- Boosted accuracy from 50% to 90% to enable tech preview and the first IBM Power GA with Spyre. Delivered prompt optimizations leading to significant accuracy improvements, and upskilled the IBM Power team to complete tech preview AI solutions, paving the way for Oct 2025 GA and OpenShift AI on Power with Spyre for 1Q 2026.
- Reduced IBM Consulting Advantage's (ICA) reliance on large third-party models by consolidating fragmented assistants onto Granite 3 8B. Onboarded Granite to the ICA platform and migrated top workloads, driving Granite adoption from 0% (Sep 2024) to 47% (Jul 2025).
- Prevented Verizon from canceling its GenAI contracts on Watsonx.ai platform and directly contributed to ~$3M in new revenue by addressing urgent legal constraints tied to a Meta-related EU issue and improving Granite performance by 50%+, making it the only model Verizon could legally and technically approve.
- Created an automated tool generating common API commands for Granite models informed by recurring usage and debugging patterns observed across real customer enterprise use cases.
- Delivered executive-level briefings, technical progress reports, and strategic insights driving leadership alignment across research, software, and hardware divisions.
CARROT LLM Router & Spyre Simulator
Experience the Pareto-optimal routing algorithms published in Állysson's CARROT paper and deployed within IBM Watsonx.ai and the IBM Consulting Gateway.
Patents & Research Publications
Spanning frontier foundation model routing, edge computer vision for sustainable agriculture, biometric access security, and adaptive cognitive interfaces.
Access control using multi-authentication factors
Novel cognitive system and methodology for dynamically orchestrating multi-factor authentication policies based on continuous environmental context and risk scoring.
Method and system for adjusting digital content composition rules based on physiological and unintended signals
Adaptive cognitive UI architecture that continuously recalibrates layout, cognitive load, and multimedia pacing by interpreting biometric and involuntary physiological feedback.
Sistema de gerenciamento e método de identificação de animais
Automated biometric tracking and sensor identification telemetry system engineered for distributed livestock inventory and veterinary health analytics.
Method and system for reporting parking space availability
Distributed sensor mesh and machine learning telemetry mechanism that predicts, detects, and broadcasts urban parking spot occupancy in real time.
Access control using multi-authentication factors (Continuation / Enhanced Methods)
Advanced cryptographic access framework utilizing behavioral entropy and distributed verification nodes for zero-trust authorization.
Portfolio & Future Labs
Production architectures deployed to millions of enterprise users alongside active explorations in next-generation autonomous agent grids and edge AI silicon.
CARROT — Cost-Aware Rate-Optimal LLM Router
Productionized routing engine deployed inside Watsonx.ai that dynamically evaluates prompt complexity, model SLA requirements, and token budgets to route queries between small distilled models (Granite 3 8B) and frontier large models with minimal latency penalty.
WXA4z Agent Co-Design on Spyre AIU
Deep silicon-software co-optimization enabling multi-step agentic AI workflows on IBM's specialized AI accelerator (Spyre AIU) for System Z and Power architectures. Pushed accuracy from sub-20% to ~90% parity with top frontier open models.
IBM Consulting Advantage (ICA) Granite Unification
Strategic migration platform that consolidated fragmented multi-vendor assistants onto Granite 3 8B, providing zero-data-leakage compliance, dramatic cost containment, and verified latency guarantees.
IBM Data Pile (Pretraining Pipeline Engine)
The core data ingestion and multiprocessing crawler engine powering IBM Data Pile (formerly Blue Pile), the proprietary pre-training corpus for IBM's Granite models. Engineered parallelized pipelines reducing crawler completion cycles drastically.
Autonomous Agent Orchestration Grid (Future Lab)
Researching decentralized agent communication protocols with formal state verification, dynamic tool synthesis, and hardware-aware resource allocation across hybrid cloud and on-prem mainframes.
Sub-Watt Agentic Runtime for Edge Silicon (Future Lab)
Exploring the boundary between microcontrollers, low-power NPU edge accelerators, and local agentic decision loops without cloud dependency.
Expanding the Frontier of Autonomous Multi-Agent Systems
New open-source agent frameworks, neuro-symbolic reasoning runtimes, and benchmark datasets are continually in development at IBM Research. Future projects will be mirrored here with live interactive sandboxes.
Technical Capabilities & Credentials
Bridging theoretical foundation model research with low-level kernel architectures, hardware accelerators, and telecommunications standards.
Generative & Agentic AI
AI Hardware & Silicon Acceleration
Data Engineering & Systems Architecture
Telecom, Standards & Mobile Systems
Connect & Collaborate
Available for executive briefings, keynote lectures, research partnerships, and strategic consultations on hardware-accelerated agentic AI systems.