Aisulu Tustykbayeva | Senior TPM, Cloud & AI Programs
I build the systems cloud and AI teams run on.
Open to platform, infrastructure, and AI TPM roles in the Bay Area. Email me
You hire a TPM when a program has no owner, no trusted numbers, and a date that cannot move. That is the work I do.
No cadence. No metrics you trust. I build both from zero, so risk shows up early, in a form you can act on.
I map the block, name the owner, and close it. No direct reports. No mandate. The system has to work anyway.
Compliance run as a program, not a scramble: scope, evidence, and gates across every team.
Zero-downtime design, with checks at each step and no margin for data loss.
Provisioning built as a product: automation, gates, and a rollout teams adopt without a mandate.
Expansion run across vendors, sites, and countries, tied to the sales pipeline it unblocks.
Migrations and expansion that cannot stop the business. Kubernetes, Terraform, GitOps, OpenStack, GraphQL, Kafka.
Audits that gate revenue, run as programs, not scrambles. PCI DSS 4.0 across 7 data centers.
For leaders with no numbers they trust. Metrics from zero: SQL, dashboards, forecasts.
For AI features that must be safe to ship. Evals, release gates, model routing. Kai: a LoRA fine-tune plus a safety harness, live in public.
PS Cloud Services: a multi-tenant IaaS platform across 7 data centers in Kazakhstan and Uzbekistan, with the API gateway serving 1.2M aggregate requests/second at peak. Enterprise customers include fintech and oil production operators with contractual SLA and compliance requirements. Engineering organized across 8 service teams. None reported to me. I ran programs across all 8.
Built the company's first TPM practice from zero. Defined the operating model, established program governance, and grew the function over 4.5 years. Two direct reports promoted within 18 months.
| Program | Outcome | The hard constraint | Teams spanned | Skills applied |
|---|---|---|---|---|
| K8s provisioning automation (Terraform + GitOps) | 96% provisioning time reduction (4–6 hrs → 10–15 min). Error rate down 10×. | 8 skeptics. One never converted. System works regardless. | DevOps, K8s Platform, DBaaS, Networking, Security |
Hard
Kubernetes / Helm
Terraform state management
GitOps CI gates
TOCTOU risk mitigation
Soft
Change management
Adoption without mandate
Rollout sequencing by trust
|
| CRM migration via GraphQL gateway | Zero downtime, $20M billing ARR protected. 12 months to 4. Query p99: 3–5s → <500ms. | Live $20M ARR billing base. No margin for data loss. 5 platforms in disagreement on sequencing. | Platform, DBaaS, K8s, LBaaS, SRE |
Hard
GraphQL gateway
Kafka — change stream and cache invalidation
Dual-write migration
Data validation pipelines
PostgreSQL / WHMCS
Soft
Decision documentation
Root cause reframing
Schedule pressure resistance
|
| PCI DSS 4.0 certification, 7 DCs | Certified 3 weeks early. $3.2M in contracts signed by customers because we passed the audit. Deploy frequency 40+/week maintained. | Competitor on the same deadline. 8 service teams, no direct authority. | LBaaS, DBaaS, K8s, Object Storage, Networking, Security, SRE, DevOps |
Hard
PCI DSS 4.0
RAG evidence dashboards
Soft
QSA negotiation
Concession strategy
Amber rule governance
Influence without authority
|
| Delivery operating system | On-time delivery: 40% to 85%, company-wide. | No mandate. Adoption had to be earned, not assigned. | All 8 service teams + Product, Legal, Finance |
Hard
OKR / KPI framework design
RAID logging
Dependency tracking
Operating model design
Soft
Influence without authority
Earned adoption
Company-wide change management
|
| DC expansion + 100 Gbps optical ring | 4 to 7 sites, 99.95% SLA maintained. $8M in deals unblocked. Sub-50ms failover. | 6 vendors, 2 countries. Design gap surfaced mid-program in synthetic load test. | Networking, Platform, SRE, Security + 6 external vendors |
Hard
DWDM optical networking
G.8032 ERPS
BFD + BGP
Multi-DC architecture
etcd quorum topology
Soft
Vendor negotiation
Physical site verification
Financial modeling for technical decisions
|
| Cross-team observability & deadlock resolution | 6-month deadlock resolved in 8 weeks. $2M+ pricing decision enabled. | No owner. Previous mandate produced silent resistance. Neither team would move first. | Product, Infrastructure (LBaaS, K8s, Compute, Object Storage), Data |
Hard
Prometheus / Grafana
Metrics taxonomy design
PoC design & execution
Schema freeze governance
Soft
Resistance root-cause mapping
Rollout sequencing by trust
Ambiguity ownership
Program durability under leadership change
|
| Infrastructure capacity planning & demand modeling | Forecast variance ±35%→±21% (40% improvement). Bottlenecks identified ahead of each DC expansion phase. | No demand telemetry. Forecasts based on sales pipeline estimates only. | Infrastructure, Finance, Sales, Engineering |
Hard
What-if simulation design
Cost modeling
Traffic forecasting
Demand telemetry
Soft
Cross-functional data gathering
Forecast credibility building
Executive communication
|
4–6 hours to 10–15 min. Error rate down 10×.
4–6 weeks to 1 week via GraphQL gateway across 5 legacy systems.
Company-wide. Up from 40%.
A safe AI tutor for children 3–8, co-built with my wife. Two layers. Layer 1: a Gemma 4 E2B model fine-tuned with QLoRA on a purpose-built 564-example dataset — age cohorts 3–4 / 5–6 / 7–8 with sentence-length caps, literal-mode and Spanish/Kazakh code-switching variants, 96 safety-refusal examples — exported as a 4-bit GGUF for low-cost tablets. Layer 2: a deployed FastAPI harness — Presidio PII scrubbing before inference, retrieval over a pre-vetted knowledge base, a pre-flight safety judge that checks every response before it streams, and Prometheus metrics for safety blocks and reading level.
Why it matters: In regulated AI, the harness and the evals are the product; the model is a component. Kai's safety rules live in a written constitution, enforced as deterministic acceptance criteria — the same pattern as review gates and confidentiality boundaries in legal AI.
Stack: Python, Unsloth / TRL, FastAPI, ChromaDB, Presidio, Prometheus · COPPA / EU AI Act research · Submitted to the Stanford Accelerator for Learning
Model card →Slurm job failure classifier for a GPU cluster (simulated cluster). Catches node-level faults (dead GPUs, driver failures, thermal throttling) before they cascade into downstream job failures. Stops scheduling to affected nodes. Outputs P1-to-P4 SLA-tiered alerts, runbooks per failure class, and cost-attributed reports per node.
Why it matters: A dead node that keeps accepting jobs burns cluster hours and delays training runs. The right intervention is at the scheduler, not the postmortem. Same framing a fleet ops team uses to make decommission decisions.
Stack: Python, Slurm, Prometheus, Grafana · NVIDIA NCA-AIIO (earned Sep 2026)
Code →Interactive tool for structured LLM prompt optimization. Engineers routinely cargo-cult DSPy settings without understanding what GEPA/BayesOpt is actually optimizing or why the labeled dataset size matters. Built a 5-step walkthrough that enforces data minimums, parses plain English into typed Signatures, runs optimization with live progress, and shows before/after metric comparison.
Design constraint: GEPA requires a minimum labeled dataset before optimization produces meaningful results. Running without that gate produces confidently wrong output, the same failure class as building migration resolvers before completing the identity map. The gate is not optional; it's surfaced at step 2, not buried in docs.
TPM framing: Mirrors wave-based migration planning. Each phase has a defined entry condition, a transformation, and an exit criterion. Same gate-and-validate structure, different domain.
Code →Consulting as TPM at Algebras AI since March 2026. In parallel, building depth in GPU and AI infrastructure: a GPU fleet failure classifier (Python, Slurm, Prometheus), and a prompt optimization tool for ML engineers (Streamlit, DSPy). Both linked in Projects.
Open to: platform & infrastructure TPM, AI TPM, capacity & forecasting roles — San Francisco Bay Area.
My first job was in sales. I noticed the enrollment process was broken, told the CEO, and ran the fix. I've been doing that ever since. Built the TPM function from scratch, which meant owning whatever the company needed: networking, compliance, migrations, cost governance. Often at once.
What I'm good at: finding the unlabeled problem that costs money, getting deep enough to act on it, and not becoming the engineer in the process.
San Francisco, CA · Green Card holder
During the COVID-19 response in Astana, organized delivery of 4,500+ hot meals to frontline medical workers across multiple hospitals during lockdown, from idea to execution in 72 hours. Received a formal letter of gratitude from the Minister of Health of Kazakhstan.
Coverage: Forbes Kazakhstan · New Times People of Year
Funded peer-reviewed air quality research in Almaty in 2021. Findings published in Atmosphere (MDPI): Seasonal and Spatial Variation of Volatile Organic Compounds in Ambient Air of Almaty City, Kazakhstan.
Founded Anti.Smog in 2019: air pollution masks tied to a research fundraising campaign. Created the first Kazakhstani brand of weighted blankets in 2021.